diff --git a/docker/npu_patch/miles.patch b/docker/npu_patch/miles.patch
index 9b9a02d61d0..0708d4d6717 100644
--- a/docker/npu_patch/miles.patch
+++ b/docker/npu_patch/miles.patch
@@ -340,7 +340,7 @@ diff --git a/miles/utils/external_utils/command_utils.py b/miles/utils/external_
 index d016e01ac..08b4d6eff 100644
 --- a/miles/utils/external_utils/command_utils.py
 +++ b/miles/utils/external_utils/command_utils.py
-@@ -193,6 +193,112 @@ def execute_train(
+@@ -193,6 +193,107 @@ def execute_train(
          )
  
  
@@ -434,18 +434,13 @@ index d016e01ac..08b4d6eff 100644
 +    )
 +
 +    if get_bool_env_var("SLIME_SCRIPT_ENABLE_RAY_SUBMIT", "1"):
-+        cmd_megatron_model_source = (
-+            f'source "{repo_base_dir}/scripts/models/{megatron_model_type}.sh" && '
-+            if megatron_model_type is not None
-+            else ""
-+        )
++        model_args = load_model_args(megatron_model_type) if megatron_model_type is not None else ""
 +        exec_command_cpu(
 +            f"export no_proxy=127.0.0.1 && export PYTHONUNBUFFERED=1 && "
-+            f"{cmd_megatron_model_source}"
 +            f'ray job submit --address="http://127.0.0.1:8265" '
 +            f"--runtime-env-json='{runtime_env_json}' "
 +            f"-- python3 {train_script} "
-+            f"{'${MODEL_ARGS[@]}' if megatron_model_type is not None else ''} "
++            f"{model_args} "
 +            f"{train_args}"
 +        )
 +
diff --git a/docs/advanced/on-policy-distillation.md b/docs/advanced/on-policy-distillation.md
index 6723abb78ac..9235c7316eb 100644
--- a/docs/advanced/on-policy-distillation.md
+++ b/docs/advanced/on-policy-distillation.md
@@ -144,7 +144,8 @@ hf download --repo-type dataset zhuzilin/dapo-math-17k --local-dir /root/dapo-ma
 
 # 2. Convert student model
 cd /root/miles
-source scripts/models/qwen3-8B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen3-8B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
     ${MODEL_ARGS[@]} \
     --hf-checkpoint /root/Qwen3-8B \
diff --git a/docs/examples/openhermes-sft.md b/docs/examples/openhermes-sft.md
index c3893afaab7..b5b7ca5bd57 100644
--- a/docs/examples/openhermes-sft.md
+++ b/docs/examples/openhermes-sft.md
@@ -27,7 +27,8 @@ If you don't already have it:
 hf download Qwen/Qwen3-4B-Base --local-dir /root/Qwen3-4B-Base
 
 cd /root/miles
-source scripts/models/qwen3-4B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen3-4B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
    --hf-checkpoint /root/Qwen3-4B-Base \
diff --git a/docs/examples/reproducibility.md b/docs/examples/reproducibility.md
index b87bd45baf6..ad1ff17b350 100644
--- a/docs/examples/reproducibility.md
+++ b/docs/examples/reproducibility.md
@@ -70,7 +70,8 @@ hf download --repo-type dataset openai/gsm8k         --local-dir /root/gsm8k
 hf download Qwen/Qwen2.5-0.5B-Instruct               --local-dir /root/Qwen2.5-0.5B-Instruct
 
 cd /root/miles
-source scripts/models/qwen2.5-0.5B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen2.5-0.5B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
    --hf-checkpoint /root/Qwen2.5-0.5B-Instruct \
diff --git a/docs/examples/search-r1.md b/docs/examples/search-r1.md
index e77a9689e57..c66ea664727 100644
--- a/docs/examples/search-r1.md
+++ b/docs/examples/search-r1.md
@@ -56,7 +56,8 @@ python $WORK_DIR/scripts/data_process/qa_search_train_merge.py \
 ```bash
 hf download Qwen/Qwen2.5-3B --local-dir /root/Qwen2.5-3B
 cd /root/miles
-source scripts/models/qwen2.5-3B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen2.5-3B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
    --hf-checkpoint /root/Qwen2.5-3B \
diff --git a/docs/getting-started/quick-start.md b/docs/getting-started/quick-start.md
index 21228399a7b..4207d4e98cd 100644
--- a/docs/getting-started/quick-start.md
+++ b/docs/getting-started/quick-start.md
@@ -65,7 +65,8 @@ map the HuggingFace weights into a sharded `torch_dist` checkpoint.
 
 ```bash
 cd /root/miles
-source scripts/models/qwen3-4B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen3-4B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 
 PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
diff --git a/docs/models/deepseek/deepseek-v4-flash.md b/docs/models/deepseek/deepseek-v4-flash.md
index e927666dc20..c1160c94d92 100644
--- a/docs/models/deepseek/deepseek-v4-flash.md
+++ b/docs/models/deepseek/deepseek-v4-flash.md
@@ -88,7 +88,8 @@ python tools/fp8_cast_bf16.py \
    --input-fp8-hf-path /root/models/DeepSeek-V4-Flash-FP8 \
    --output-bf16-hf-path /root/models/DeepSeek-V4-Flash-FP8-bf16/
 
-source scripts/models/deepseek-v4-flash.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py deepseek-v4-flash)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM torchrun \
    --nproc-per-node 4 --nnodes 8 \
    --master-addr ${MASTER_ADDR} --master-port 12345 \
diff --git a/docs/models/deepseek/deepseek.md b/docs/models/deepseek/deepseek.md
index ffc3d4cba9c..da922a7b15d 100644
--- a/docs/models/deepseek/deepseek.md
+++ b/docs/models/deepseek/deepseek.md
@@ -52,7 +52,8 @@ Then convert BF16 HF → Megatron `torch_dist`. Run on **4 separate nodes** (`NO
 
 ```bash
 cd miles/
-source scripts/models/deepseek-v3.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py deepseek-v3)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM/ torchrun \
    --nproc-per-node 8 \
    --master-addr ${MASTER_ADDR} --master-port 12345 \
diff --git a/docs/models/glm/glm4-5.md b/docs/models/glm/glm4-5.md
index 7aeb2492515..35a4e8d0458 100644
--- a/docs/models/glm/glm4-5.md
+++ b/docs/models/glm/glm4-5.md
@@ -49,7 +49,8 @@ The bash launcher does **not** convert for you — produce `$BASE_DIR/GLM-4.5-35
 
 ```bash
 cd /root/miles
-source scripts/models/glm4.5-355B-A32B.sh
+MODEL_ARGS_LINE="$(python3 scripts/model_args.py glm4.5-355B-A32B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM torchrun --nproc-per-node 8 \
    tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
diff --git a/docs/models/glm/glm4-7-flash.md b/docs/models/glm/glm4-7-flash.md
index 07162c31fa3..31b368b3f15 100644
--- a/docs/models/glm/glm4-7-flash.md
+++ b/docs/models/glm/glm4-7-flash.md
@@ -35,7 +35,8 @@ The bash launcher hardcodes `BASE_DIR=/root/shared`. The Python launcher downloa
 
 ```bash
 cd /root/miles
-source scripts/models/glm4.7-flash.sh
+MODEL_ARGS_LINE="$(python3 scripts/model_args.py glm4.7-flash)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM torchrun --nproc-per-node 8 \
    tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
diff --git a/docs/models/glm/glm4.md b/docs/models/glm/glm4.md
index 688ba67ce82..04dcd74629d 100644
--- a/docs/models/glm/glm4.md
+++ b/docs/models/glm/glm4.md
@@ -32,7 +32,8 @@ hf download --repo-type dataset zhuzilin/aime-2024     --local-dir /root/aime-20
 
 ```bash
 cd /root/miles
-source scripts/models/glm4-9B.sh
+MODEL_ARGS_LINE="$(python3 scripts/model_args.py glm4-9B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
    --hf-checkpoint /root/GLM-Z1-9B-0414 \
diff --git a/docs/models/kimi/kimi-k2.5.md b/docs/models/kimi/kimi-k2.5.md
index 5b766462c22..f4d924ff428 100644
--- a/docs/models/kimi/kimi-k2.5.md
+++ b/docs/models/kimi/kimi-k2.5.md
@@ -2,7 +2,7 @@
 title: Kimi K2.5 / K2.6
 description: Launch recipe for Kimi-K2.5, running full-parameter GRPO on 32 × 8 H200 with an INT4 actor and a BF16 reference.
 ---
-The reference launcher is [`scripts/run-kimi-k25.sh`](https://github.com/radixark/miles/blob/main/scripts/run-kimi-k25.sh), which sources the shared model definition in `scripts/models/kimi-k2-thinking.sh`.
+The reference launcher is [`scripts/run-kimi-k25.sh`](https://github.com/radixark/miles/blob/main/scripts/run-kimi-k25.sh), which loads the shared model definition from `scripts/models/kimi-k2-thinking.py`.
 
 ## 1. Model Introduction
 
@@ -79,7 +79,7 @@ ray start --address=${MASTER_ADDR}:6379 --num-gpus 8 --node-ip-address ${WORKER_
 
 ## 4. Script breakdown
 
-The launcher groups its flags into the arrays that are passed to `train.py`. The model shape comes from `MODEL_ARGS`, which is sourced from `scripts/models/kimi-k2-thinking.sh`. That definition sets the MLA latent ranks (`q_lora_rank=1536`, `kv_lora_rank=512`, `qk_head_dim=128`, `qk_pos_emb_head_dim=64`, `v_head_dim=128`), the MoE routing (384 experts, top-8, sigmoid pre-softmax scoring, FP32 router, `--moe-router-topk-scaling-factor 2.827`), and RoPE (`--rotary-base 50000`, `--rotary-scaling-factor 64.0`). The K2.5 recipe then layers the following on top:
+The launcher groups its flags into the arrays that are passed to `train.py`. The model shape comes from `MODEL_ARGS`, which is loaded from `scripts/models/kimi-k2-thinking.py`. That definition sets the MLA latent ranks (`q_lora_rank=1536`, `kv_lora_rank=512`, `qk_head_dim=128`, `qk_pos_emb_head_dim=64`, `v_head_dim=128`), the MoE routing (384 experts, top-8, sigmoid pre-softmax scoring, FP32 router, `--moe-router-topk-scaling-factor 2.827`), and RoPE (`--rotary-base 50000`, `--rotary-scaling-factor 64.0`). The K2.5 recipe then layers the following on top:
 
 - **`CKPT_ARGS`** wires up the dual checkpoint (INT4 actor via `--hf-checkpoint`, BF16 reference via `--ref-load`) together with `--megatron-to-hf-mode bridge` and `--model-name kimi_k25`.
 - **`ROLLOUT_ARGS`** and **`EVAL_ARGS`** configure GRPO sampling and periodic AIME evaluation (covered in §5.2).
diff --git a/docs/models/kimi/kimi-k2.md b/docs/models/kimi/kimi-k2.md
index 7bb410b834b..3315577492c 100644
--- a/docs/models/kimi/kimi-k2.md
+++ b/docs/models/kimi/kimi-k2.md
@@ -47,7 +47,8 @@ Convert across 4 nodes (mirror the DeepSeek-V3 procedure):
 
 ```bash
 cd /root/miles
-source scripts/models/kimi-k2.sh   # or kimi-k2-thinking.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py kimi-k2)" || exit 1   # or kimi-k2-thinking
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM/ torchrun \
    --nproc-per-node 8 \
    --master-addr ${MASTER_ADDR} --master-port 12345 \
diff --git a/docs/models/kimi/moonlight.md b/docs/models/kimi/moonlight.md
index 5f1e781ba5c..b4244127b07 100644
--- a/docs/models/kimi/moonlight.md
+++ b/docs/models/kimi/moonlight.md
@@ -33,7 +33,8 @@ hf download --repo-type dataset zhuzilin/aime-2024     --local-dir /root/aime-20
 
 ```bash
 cd /root/miles
-source scripts/models/moonlight.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py moonlight)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM torchrun --nproc-per-node 8 \
    tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
diff --git a/docs/models/mimo/mimo.md b/docs/models/mimo/mimo.md
index bec36877023..0e35bb6f458 100644
--- a/docs/models/mimo/mimo.md
+++ b/docs/models/mimo/mimo.md
@@ -33,7 +33,8 @@ hf download --repo-type dataset zhuzilin/aime-2024     --local-dir /root/aime-20
 
 ```bash
 cd /root/miles
-source scripts/models/mimo-7B-rl.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py mimo-7B-rl)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
    --hf-checkpoint /root/MiMo-7B-RL \
diff --git a/docs/models/nemotron/nemotron-3-nano-moe.md b/docs/models/nemotron/nemotron-3-nano-moe.md
index ff4efdbf4af..77baa943c58 100644
--- a/docs/models/nemotron/nemotron-3-nano-moe.md
+++ b/docs/models/nemotron/nemotron-3-nano-moe.md
@@ -125,7 +125,7 @@ memory pressure rises.
 
 ### 5.5 Notable quirks
 
-From `scripts/models/nemotron-3-nano-30b-a3b.sh` and `scripts/run-nemotron-3-nano-30b-a3b.sh`:
+From `scripts/models/nemotron-3-nano-30b-a3b.py` and `scripts/run-nemotron-3-nano-30b-a3b.sh`:
 
 - **No `--spec`**: AutoBridge + the NemotronH shim synthesize the Megatron MoE spec from HF config.
 - 128 experts, `--moe-router-topk 6`, shared expert (3712-dim).
diff --git a/docs/models/nemotron/nemotron-3-nano.md b/docs/models/nemotron/nemotron-3-nano.md
index f45af012534..75dbe1f3147 100644
--- a/docs/models/nemotron/nemotron-3-nano.md
+++ b/docs/models/nemotron/nemotron-3-nano.md
@@ -112,7 +112,7 @@ memory pressure rises.
 
 ### 5.5 Notable quirks
 
-From `scripts/models/nemotron-3-nano-4b.sh` and `scripts/run-nemotron-3-nano-4b.sh`:
+From `scripts/models/nemotron-3-nano-4b.py` and `scripts/run-nemotron-3-nano-4b.sh`:
 
 - **No `--spec`**: the AutoBridge synthesizes the Megatron spec from HF config.
 - `--position-embedding-type none` (no RoPE).
diff --git a/docs/models/qwen/qwen3-5-moe.md b/docs/models/qwen/qwen3-5-moe.md
index 0c0fd7092b5..f5f24d66c78 100644
--- a/docs/models/qwen/qwen3-5-moe.md
+++ b/docs/models/qwen/qwen3-5-moe.md
@@ -32,7 +32,8 @@ hf download --repo-type dataset zhuzilin/aime-2024     --local-dir /root/aime-20
 
 ```bash
 cd /root/miles
-source scripts/models/qwen3.5-35B-A3B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen3.5-35B-A3B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM torchrun --nproc-per-node 8 \
    tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
@@ -98,7 +99,7 @@ CPU Adam is enabled (`--optimizer-cpu-offload --overlap-cpu-optimizer-d2h-h2d --
 ### 5.5 Notable quirks
 
 - The Megatron side uses `--moe-token-dispatcher-type flex`; DeepEP isn't enabled here, unlike Qwen3-Next.
-- The model config (`scripts/models/qwen3.5-35B-A3B.sh`) reuses the Qwen3.5 spec: `--attention-output-gate`, `--rotary-base 10000000`, `--rotary-percent 0.25`, `A_log` kept in FP32 via the bridge. See [Backends Beyond Megatron](/advanced/architecture-support).
+- The model config (`scripts/models/qwen3.5-35B-A3B.py`) reuses the Qwen3.5 spec: `--attention-output-gate`, `--rotary-base 10000000`, `--rotary-percent 0.25`, `A_log` kept in FP32 via the bridge. See [Backends Beyond Megatron](/advanced/architecture-support).
 
 ## 6. Pairs Well With
 
diff --git a/docs/models/qwen/qwen3-5.md b/docs/models/qwen/qwen3-5.md
index 8169df811d9..9cfb1419c68 100644
--- a/docs/models/qwen/qwen3-5.md
+++ b/docs/models/qwen/qwen3-5.md
@@ -34,7 +34,8 @@ hf download --repo-type dataset zhuzilin/aime-2024     --local-dir /root/aime-20
 
 ```bash
 cd /root/miles
-source scripts/models/qwen3.5-4B.sh   # or qwen3.5-9B.sh / qwen3.5-27B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen3.5-4B)" || exit 1   # or qwen3.5-9B / qwen3.5-27B
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
    --hf-checkpoint /root/Qwen3.5-4B \
@@ -92,7 +93,7 @@ Only the 27 B script enables CPU Adam (`--optimizer-cpu-offload --overlap-cpu-op
 
 ### 5.5 Notable quirks
 
-From `scripts/models/qwen3.5-4B.sh` (and analogous configs for 9 B / 27 B):
+From `scripts/models/qwen3.5-4B.py` (and analogous configs for 9 B / 27 B):
 
 - `--spec miles_plugins.models.qwen3_5 get_qwen3_5_spec` — attention-output gate, `A_log` parameter handling.
 - `--rotary-base 10000000`, `--rotary-percent 0.25`.
diff --git a/docs/models/qwen/qwen3-6-moe.md b/docs/models/qwen/qwen3-6-moe.md
index 1e0b21c87d9..d8b988d49cd 100644
--- a/docs/models/qwen/qwen3-6-moe.md
+++ b/docs/models/qwen/qwen3-6-moe.md
@@ -47,7 +47,8 @@ hf download Qwen/Qwen3.6-35B-A3B --local-dir /root/models/Qwen3.6-35B-A3B
 
 ```bash
 cd /root/miles
-source scripts/models/qwen3.6-35B-A3B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen3.6-35B-A3B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM torchrun --nproc-per-node 8 \
    tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
@@ -136,7 +137,7 @@ CPU Adam is enabled (`--optimizer-cpu-offload --overlap-cpu-optimizer-d2h-h2d --
 
 ### 5.5 Notable quirks
 
-From `scripts/models/qwen3.6-35B-A3B.sh` and `scripts/run_qwen3_6_35b_a3b_mtp.py`:
+From `scripts/models/qwen3.6-35B-A3B.py` and `scripts/run_qwen3_6_35b_a3b_mtp.py`:
 
 - `--spec miles_plugins.models.qwen3_5 get_qwen3_5_spec` — Qwen3.6 reuses the Qwen3.5 spec.
 - 256 experts, `--moe-router-topk 8`, `--moe-router-score-function softmax`.
diff --git a/docs/models/qwen/qwen3-6.md b/docs/models/qwen/qwen3-6.md
index adac7cbba70..9829ed746a7 100644
--- a/docs/models/qwen/qwen3-6.md
+++ b/docs/models/qwen/qwen3-6.md
@@ -45,7 +45,8 @@ hf download --repo-type dataset zhuzilin/aime-2024     --local-dir /root/aime-20
 
 ```bash
 cd /root/miles
-source scripts/models/qwen3.6-27B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen3.6-27B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
    --hf-checkpoint /root/Qwen3.6-27B \
@@ -109,7 +110,7 @@ CPU Adam is enabled (`--optimizer-cpu-offload --overlap-cpu-optimizer-d2h-h2d --
 
 ### 5.5 Notable quirks
 
-From `scripts/models/qwen3.6-27B.sh`:
+From `scripts/models/qwen3.6-27B.py`:
 
 - `--spec miles_plugins.models.qwen3_5 get_qwen3_5_spec` — Qwen3.6 reuses the Qwen3.5 spec (gated attention, FP32 `A_log`).
 - `--rotary-base 10000000`, `--rotary-percent 0.25`.
diff --git a/docs/models/qwen/qwen3-moe.md b/docs/models/qwen/qwen3-moe.md
index 4c8f92c2b42..f043fc6ecca 100644
--- a/docs/models/qwen/qwen3-moe.md
+++ b/docs/models/qwen/qwen3-moe.md
@@ -47,7 +47,8 @@ hf download Qwen/Qwen3-235B-A22B-FP8 --local-dir $BASE_FOLDER/Qwen3-235B-A22B-FP
 ### 3.3 HF → Megatron `torch_dist` conversion
 
 ```bash
-source scripts/models/qwen3-30B-A3B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen3-30B-A3B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM torchrun --nproc-per-node 8 \
    tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
diff --git a/docs/models/qwen/qwen3-next.md b/docs/models/qwen/qwen3-next.md
index 3aff187dbe9..bc050a9073f 100644
--- a/docs/models/qwen/qwen3-next.md
+++ b/docs/models/qwen/qwen3-next.md
@@ -42,7 +42,8 @@ hf download --repo-type dataset zhuzilin/aime-2024     --local-dir $BASE_FOLDER/
 
 ```bash
 cd /root/miles
-source scripts/models/qwen3-next-80B-A3B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen3-next-80B-A3B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM torchrun --nproc-per-node 8 \
    tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
diff --git a/docs/models/qwen/qwen3.md b/docs/models/qwen/qwen3.md
index b2f07f0e839..03f410a3488 100644
--- a/docs/models/qwen/qwen3.md
+++ b/docs/models/qwen/qwen3.md
@@ -39,7 +39,8 @@ hf download --repo-type dataset zhuzilin/aime-2024     --local-dir /root/aime-20
 
 ```bash
 cd /root/miles
-source scripts/models/qwen3-4B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen3-4B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
    --hf-checkpoint /root/Qwen3-4B \
@@ -57,9 +58,9 @@ cd /root/miles
 bash scripts/run-qwen3-4B.sh
 ```
 
-Other variants follow the same pattern — replace the script name (`run-qwen3-32B.sh`, etc.) and the `qwen3-XB.sh` model config.
+Other variants follow the same pattern — replace the script name (`run-qwen3-32B.sh`, etc.) and the `qwen3-XB.py` model config.
 
-The Qwen3-4B-Instruct-2507 config (`scripts/models/qwen3-4B-Instruct-2507.sh`) just sets `MODEL_ARGS_ROTARY_BASE=5000000` and re-sources `qwen3-4B.sh` — source it when converting / launching the Instruct-2507 checkpoint.
+The Qwen3-4B-Instruct-2507 config (`scripts/models/qwen3-4B-Instruct-2507.py`) just calls `qwen3-4B` with `rotary_base=5000000` (`MODEL_ARGS_ROTARY_BASE` still works as an environment override) — load it when converting / launching the Instruct-2507 checkpoint.
 
 ## 5. Recipe Configuration
 
diff --git a/docs/models/thinkingmachines/inkling-small.md b/docs/models/thinkingmachines/inkling-small.md
index 5c0f5bc1aea..398ac67fa01 100644
--- a/docs/models/thinkingmachines/inkling-small.md
+++ b/docs/models/thinkingmachines/inkling-small.md
@@ -39,7 +39,7 @@ python scripts/run_inkling.py train \
    --sglang-context-length 4096 --rollout-max-response-len 2048
 ```
 
-The model definition lives in `scripts/models/inkling-small.sh` (`MODEL_ARGS_NUM_LAYERS` overrides the layer count for sliced smoke/parity checkpoints). HF → `torch_dist` conversion uses the same tool as Inkling with this recipe file — a single 8-GPU node (TP8 EP8) converts it in one pass.
+The model definition lives in `scripts/models/inkling-small.py` (`MODEL_ARGS_NUM_LAYERS` overrides the layer count for sliced smoke/parity checkpoints). HF → `torch_dist` conversion uses the same tool as Inkling with this recipe file — a single 8-GPU node (TP8 EP8) converts it in one pass.
 
 ## 4. Validated parallelism
 
diff --git a/docs/models/thinkingmachines/inkling.md b/docs/models/thinkingmachines/inkling.md
index 31416e01328..801f9653f00 100644
--- a/docs/models/thinkingmachines/inkling.md
+++ b/docs/models/thinkingmachines/inkling.md
@@ -72,11 +72,12 @@ Pass `--hf-checkpoint <path>` to the launcher when the weights are already on a
 
 ### 4.2 HF → Megatron `torch_dist` conversion
 
-Inkling ships in BF16, so conversion is a single distributed `torch_dist` shard (no precision cast). The model definition comes from `scripts/models/inkling.sh`:
+Inkling ships in BF16, so conversion is a single distributed `torch_dist` shard (no precision cast). The model definition comes from `scripts/models/inkling.py`:
 
 ```bash
 cd /root/miles
-source scripts/models/inkling.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py inkling)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CONVERT_KEEP_PP1=1 PYTHONPATH=/root/Megatron-LM torchrun \
    --nproc-per-node 4 --nnodes 4 \
    --master-addr ${MASTER_ADDR} --master-port 12345 \
diff --git a/docs/platforms/amd.md b/docs/platforms/amd.md
index f1cd7bd3806..07f9baa2c8c 100644
--- a/docs/platforms/amd.md
+++ b/docs/platforms/amd.md
@@ -68,7 +68,8 @@ ROCm converter is in development.
 
 ```bash
 cd /root/miles
-source scripts/models/qwen3-4B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen3-4B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 MEGATRON_LM_PATH=$(pip list | grep megatron-core | awk '{print $NF}')
 
 PYTHONPATH=${MEGATRON_LM_PATH} python tools/convert_hf_to_torch_dist.py \
diff --git a/docs/user-guide/argument-groups.md b/docs/user-guide/argument-groups.md
index 11e52e7e6b8..bb20a53b766 100644
--- a/docs/user-guide/argument-groups.md
+++ b/docs/user-guide/argument-groups.md
@@ -11,7 +11,7 @@ when you need the full default and type for an individual flag.
 
 | Group | Owns | Typical source |
 |---|---|---|
-| [`MODEL_ARGS`](#model-args) | Architecture constants and plugin specs | `scripts/models/<family>.sh` |
+| [`MODEL_ARGS`](#model-args) | Architecture constants and plugin specs | `scripts/models/<family>.py` |
 | [`CKPT_ARGS`](#ckpt-args) | Actor, reference, HF tokenizer/config, save paths | Launch script |
 | [`ROLLOUT_ARGS`](#rollout-args) | Prompt data, sampling, reward, train/eval batch flow | Launch script |
 | [`EVAL_ARGS`](#eval-args) | Evaluation datasets and eval-only sampling overrides | Launch script |
@@ -24,7 +24,7 @@ when you need the full default and type for an individual flag.
 ## MODEL_ARGS - architecture constants
 
 `MODEL_ARGS` tells Megatron what model it is instantiating. Megatron cannot infer all
-architecture details from a HuggingFace checkpoint, so each recipe sources a matching
+architecture details from a HuggingFace checkpoint, so each recipe loads a matching
 file from `scripts/models/`.
 
 Common entries:
diff --git a/docs/user-guide/training-script-walkthrough.md b/docs/user-guide/training-script-walkthrough.md
index 341aef2dcc8..6499e630ae0 100644
--- a/docs/user-guide/training-script-walkthrough.md
+++ b/docs/user-guide/training-script-walkthrough.md
@@ -29,15 +29,16 @@ off to `train.py`:
 ## MODEL_ARGS — architecture constants
 
 Megatron needs the model architecture hardcoded at launch because it cannot introspect
-a HuggingFace checkpoint. Miles therefore sources a matching bash file from
-`scripts/models/<family>.sh`:
+a HuggingFace checkpoint. Miles therefore loads a matching python file from
+`scripts/models/<family>.py`:
 
 ```bash
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/glm4-9B.sh"
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" glm4-9B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 ```
 
-The sourced file sets `MODEL_ARGS=(--num-layers ... --hidden-size ... --rotary-base ...)`.
+The loaded file prints `--num-layers ... --hidden-size ... --rotary-base ...` on one line.
 
 <Warning>
 
@@ -46,7 +47,8 @@ padding, or normalization epsilon. Diff the `config.json` against the file in
 `scripts/models/` before you run, and override anything that drifts:
 
 ```bash
-source "${SCRIPT_DIR}/models/glm4-9B.sh"
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" glm4-9B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 MODEL_ARGS+=(--rotary-base 10000)
 ```
 
diff --git a/docs/user-guide/usage.md b/docs/user-guide/usage.md
index 7dccf4d29f6..359dee8655a 100644
--- a/docs/user-guide/usage.md
+++ b/docs/user-guide/usage.md
@@ -106,7 +106,8 @@ the next run. Requires `--save` to be set.
 ### HuggingFace → torch_dist
 
 ```bash
-source scripts/models/<family>.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py <family>)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
    --hf-checkpoint /root/<model> \
diff --git a/examples/experimental/eval/nemo_skills/README.md b/examples/experimental/eval/nemo_skills/README.md
index e2a3aaad703..c9d5bf79f48 100644
--- a/examples/experimental/eval/nemo_skills/README.md
+++ b/examples/experimental/eval/nemo_skills/README.md
@@ -138,7 +138,8 @@ You need to convert the HF model to the format required by Megatron-LM. Ensure y
 
 ```bash
 # Source model arguments
-source scripts/models/qwen3-4B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen3-4B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 
 # Convert model
 PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
diff --git a/examples/experimental/eval/scripts/run-qwen3-32B.sh b/examples/experimental/eval/scripts/run-qwen3-32B.sh
index 525bfe357d5..cd09054754a 100644
--- a/examples/experimental/eval/scripts/run-qwen3-32B.sh
+++ b/examples/experimental/eval/scripts/run-qwen3-32B.sh
@@ -30,7 +30,8 @@ echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
 REPO_ROOT="$(cd -- "${SCRIPT_DIR}/../../../.." &>/dev/null && pwd)"
-source "${REPO_ROOT}/scripts/models/qwen3-32B.sh"
+MODEL_ARGS_LINE="$(python3 "${REPO_ROOT}/miles/utils/external_utils/model_args_utils.py" "qwen3-32B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 
 # Store eval/delegate settings in a YAML config similar to examples/experimental/eval_multi_task.
 EVAL_CONFIG_PATH=${SKILLS_EVAL_CONFIG_PATH:-"${REPO_ROOT}/examples/experimental/eval/scripts/multi_tasks.yaml"}
diff --git a/examples/experimental/eval/scripts/run-qwen3-4B.sh b/examples/experimental/eval/scripts/run-qwen3-4B.sh
index e2647973c80..250e6c996a8 100644
--- a/examples/experimental/eval/scripts/run-qwen3-4B.sh
+++ b/examples/experimental/eval/scripts/run-qwen3-4B.sh
@@ -31,7 +31,8 @@ echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
 REPO_ROOT="$(cd -- "${SCRIPT_DIR}/../../../.." &>/dev/null && pwd)"
-source "${REPO_ROOT}/scripts/models/qwen3-4B.sh"
+MODEL_ARGS_LINE="$(python3 "${REPO_ROOT}/miles/utils/external_utils/model_args_utils.py" "qwen3-4B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 
 # Store eval/delegate settings in a YAML config similar to examples/experimental/eval_multi_task.
 EVAL_CONFIG_PATH=${SKILLS_EVAL_CONFIG_PATH:-"${REPO_ROOT}/examples/experimental/eval/scripts/multi_tasks.yaml"}
diff --git a/examples/experimental/eval_multi_task/multi_task.sh b/examples/experimental/eval_multi_task/multi_task.sh
index 090c461a005..6a5aeca0bb8 100644
--- a/examples/experimental/eval_multi_task/multi_task.sh
+++ b/examples/experimental/eval_multi_task/multi_task.sh
@@ -25,7 +25,8 @@ echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
 REPO_ROOT="$(cd -- "${SCRIPT_DIR}/../../.." &>/dev/null && pwd)"
-source "${REPO_ROOT}/scripts/models/qwen3-4B.sh"
+MODEL_ARGS_LINE="$(python3 "${REPO_ROOT}/miles/utils/external_utils/model_args_utils.py" "qwen3-4B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 EVAL_CONFIG_PATH="${REPO_ROOT}/examples/experimental/eval_multi_task/multi_task.yaml"
 
 CKPT_ARGS=(
diff --git a/examples/experimental/formal_math/single_round/run_minimal.py b/examples/experimental/formal_math/single_round/run_minimal.py
index fa19ccf4524..7e7d7745387 100644
--- a/examples/experimental/formal_math/single_round/run_minimal.py
+++ b/examples/experimental/formal_math/single_round/run_minimal.py
@@ -8,6 +8,8 @@
 import subprocess
 from pathlib import Path
 
+from miles.utils.external_utils.model_args_utils import load_model_args
+
 repo_base_dir = Path(os.path.abspath(__file__)).resolve().parents[4]
 
 MODEL_NAME, MODEL_TYPE = "Qwen3-8B", "qwen3-8B"
@@ -131,11 +133,10 @@
 
 cmd = (
     f"export PYTHONUNBUFFERED=1 && "
-    f'source "{repo_base_dir}/scripts/models/{MODEL_TYPE}.sh" && '
     f'ray job submit --address="http://127.0.0.1:8265" '
     f"--runtime-env-json='{runtime_env_json}' "
     f"-- python3 train.py "
-    "${MODEL_ARGS[@]} "
+    f"{load_model_args(MODEL_TYPE)} "
     f"{train_args}"
 )
 
diff --git a/examples/experimental/multi_agent/run-qwen3-30B-A3B-multi-agent.sh b/examples/experimental/multi_agent/run-qwen3-30B-A3B-multi-agent.sh
index fa484f0468d..3ca29aaefce 100644
--- a/examples/experimental/multi_agent/run-qwen3-30B-A3B-multi-agent.sh
+++ b/examples/experimental/multi_agent/run-qwen3-30B-A3B-multi-agent.sh
@@ -24,8 +24,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/../../../scripts/models/qwen3-30B-A3B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../../../miles/utils/external_utils/model_args_utils.py" "qwen3-30B-A3B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint /root/Qwen3-30B-A3B
    #--hf-checkpoint /root/Qwen3-30B-A3B-FP8
diff --git a/examples/experimental/reproducibility/README.md b/examples/experimental/reproducibility/README.md
index 08759d7aa16..cedfc7d366e 100644
--- a/examples/experimental/reproducibility/README.md
+++ b/examples/experimental/reproducibility/README.md
@@ -34,7 +34,8 @@ hf download Qwen/Qwen2.5-0.5B-Instruct --local-dir /root/Qwen2.5-0.5B-Instruct
 
 # convert ckpt
 cd miles/
-source scripts/models/qwen2.5-0.5B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen2.5-0.5B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM/ python \
    tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
diff --git a/examples/experimental/reproducibility/run-qwen2.5-0.5B-gsm8k.sh b/examples/experimental/reproducibility/run-qwen2.5-0.5B-gsm8k.sh
index f8d498ee2ee..b84b66c243b 100644
--- a/examples/experimental/reproducibility/run-qwen2.5-0.5B-gsm8k.sh
+++ b/examples/experimental/reproducibility/run-qwen2.5-0.5B-gsm8k.sh
@@ -16,7 +16,8 @@ set -ex
 export PYTHONUNBUFFERED=1
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/../../../scripts/models/qwen2.5-0.5B.sh"
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../../../miles/utils/external_utils/model_args_utils.py" "qwen2.5-0.5B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 
 CKPT_ARGS=(
    --hf-checkpoint /root/Qwen2.5-0.5B-Instruct/
diff --git a/examples/experimental/search-r1/README.md b/examples/experimental/search-r1/README.md
index 4cab49e363d..0089e021e5c 100644
--- a/examples/experimental/search-r1/README.md
+++ b/examples/experimental/search-r1/README.md
@@ -51,7 +51,8 @@ hf download Qwen/Qwen2.5-3B --local-dir /root/Qwen2.5-3B
 
 # mcore checkpoint
 cd /root/miles
-source scripts/models/qwen2.5-3B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen2.5-3B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
     ${MODEL_ARGS[@]} \
     --hf-checkpoint /root/Qwen2.5-3B \
@@ -70,17 +71,14 @@ SEARCH_R1_CONFIGS = {
     "max_turns": 2,
     "topk": 3,
     "search_concurrency": 256,
-
     # ============== Search Backend Selection ==============
     "search_backend": "local",  # Options: "local" or "google"
-
     # ============== Local Search Configuration ==============
     # (Only used when search_backend="local")
     "local": {
         "search_url": "http://127.0.0.1:8000/retrieve",  # URL of your local retrieval server
         "proxy": None,
     },
-
     # ============== Google Search Configuration ==============
     # (Only used when search_backend="google")
     "google": {
@@ -88,10 +86,8 @@ SEARCH_R1_CONFIGS = {
         "snippet_only": True,
         "proxy": None,
     },
-
     # ============== Log Probability Collection ==============
     "return_logprob": True,  # Set to True to collect log probabilities (required for TIS)
-
     # ============== Reward Model Configuration ==============
     "format_score": 0.2,
 }
diff --git a/examples/experimental/search-r1/run_qwen2.5_3B.sh b/examples/experimental/search-r1/run_qwen2.5_3B.sh
index 798a29d75ba..ab0dd75cb3c 100644
--- a/examples/experimental/search-r1/run_qwen2.5_3B.sh
+++ b/examples/experimental/search-r1/run_qwen2.5_3B.sh
@@ -16,7 +16,8 @@ set -ex
 export PYTHONUNBUFFERED=1
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/../../../scripts/models/qwen2.5-3B.sh"
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../../../miles/utils/external_utils/model_args_utils.py" "qwen2.5-3B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 
 CKPT_ARGS=(
    --hf-checkpoint /root/Qwen2.5-3B/
diff --git a/examples/experimental/strands_sglang/README.md b/examples/experimental/strands_sglang/README.md
index 6fe4cce4d47..158b113d011 100644
--- a/examples/experimental/strands_sglang/README.md
+++ b/examples/experimental/strands_sglang/README.md
@@ -36,7 +36,8 @@ hf download Qwen/Qwen3-8B --local-dir /root/models/Qwen/Qwen3-8B
 
 # mcore checkpoint
 cd /root/miles
-source scripts/models/qwen3-8B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen3-8B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
     ${MODEL_ARGS[@]} \
     --hf-checkpoint /root/models/Qwen/Qwen3-8B \
@@ -49,6 +50,7 @@ Following [Retool](https://arxiv.org/abs/2504.11536), we use `dapo-math-17k` as
 
 ```python
 from datasets import load_dataset
+
 ds = load_dataset("zhuzilin/dapo-math-17k", split="train")
 ds.to_json("/root/data/dapo-math-17k.jsonl", orient="records", lines=True)
 ```
@@ -57,6 +59,7 @@ and `aime-2024` as eval data:
 
 ```python
 from datasets import load_dataset
+
 ds = load_dataset("zhuzilin/aime-2024", split="train")
 ds.to_json("/root/data/aime-2024.jsonl", orient="records", lines=True)
 ```
diff --git a/examples/experimental/strands_sglang/strands_qwen3_8b.sh b/examples/experimental/strands_sglang/strands_qwen3_8b.sh
index 9e4aa1f45e4..79a49a6467a 100644
--- a/examples/experimental/strands_sglang/strands_qwen3_8b.sh
+++ b/examples/experimental/strands_sglang/strands_qwen3_8b.sh
@@ -28,8 +28,8 @@ echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
 MILES_ROOT="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/../.." &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/../../../scripts/models/qwen3-8B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../../../miles/utils/external_utils/model_args_utils.py" "qwen3-8B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 # Generate timestamp suffix for save path
 TIMESTAMP_SUFFIX=$(date +%Y%m%d_%H%M%S)
 
diff --git a/examples/experimental/tau-bench/README.md b/examples/experimental/tau-bench/README.md
index 2959ef1dd7f..b0179ad7017 100644
--- a/examples/experimental/tau-bench/README.md
+++ b/examples/experimental/tau-bench/README.md
@@ -33,7 +33,8 @@ hf download Qwen/Qwen3-4B-Instruct-2507 --local-dir /root/Qwen3-4B-Instruct-2507
 
 # mcore checkpoint
 cd /root/miles
-source scripts/models/qwen3-4B-Instruct-2507.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen3-4B-Instruct-2507)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
     ${MODEL_ARGS[@]} \
     --hf-checkpoint /root/Qwen3-4B-Instruct-2507 \
diff --git a/examples/experimental/tau-bench/run_qwen3_4B.sh b/examples/experimental/tau-bench/run_qwen3_4B.sh
index 172834e79f1..184c0fa5e9d 100644
--- a/examples/experimental/tau-bench/run_qwen3_4B.sh
+++ b/examples/experimental/tau-bench/run_qwen3_4B.sh
@@ -24,7 +24,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/../../../scripts/models/qwen3-4B-Instruct-2507.sh"
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../../../miles/utils/external_utils/model_args_utils.py" "qwen3-4B-Instruct-2507")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 
 CKPT_ARGS=(
    --hf-checkpoint /root/Qwen3-4B-Instruct-2507/
diff --git a/examples/fully_async/run-qwen3-4b-fully_async.sh b/examples/fully_async/run-qwen3-4b-fully_async.sh
index 44445b86472..07dd2b6ee5f 100644
--- a/examples/fully_async/run-qwen3-4b-fully_async.sh
+++ b/examples/fully_async/run-qwen3-4b-fully_async.sh
@@ -24,8 +24,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/../../scripts/models/qwen3-4B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../../miles/utils/external_utils/model_args_utils.py" "qwen3-4B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint /root/Qwen3-4B
    #--hf-checkpoint /root/Qwen3-4B-FP8
diff --git a/examples/geo3k_vlm/run_geo3k_vlm.sh b/examples/geo3k_vlm/run_geo3k_vlm.sh
index e95c55ebb4b..3da32be58c6 100644
--- a/examples/geo3k_vlm/run_geo3k_vlm.sh
+++ b/examples/geo3k_vlm/run_geo3k_vlm.sh
@@ -188,7 +188,8 @@ else
    MILES_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/../.." &>/dev/null && pwd)"
    MODEL_ARGS_FILE=$(echo "$MODEL_NAME" | sed 's/-Instruct//g; s/-Thinking//g; s/Qwen3-VL-/qwen3-/g; s/-2B/-1.7B/g')
    # VL models require rotary-base 5000000
-   MODEL_ARGS_ROTARY_BASE=5000000 source "${MILES_DIR}/scripts/models/${MODEL_ARGS_FILE}.sh"
+   MODEL_ARGS_LINE="$(MODEL_ARGS_ROTARY_BASE=5000000 python3 "${MILES_DIR}/miles/utils/external_utils/model_args_utils.py" "${MODEL_ARGS_FILE}")" || exit 1
+   read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
    
 fi
 
diff --git a/examples/geo3k_vlm/run_geo3k_vlm_sft.sh b/examples/geo3k_vlm/run_geo3k_vlm_sft.sh
index 7975c4c07f7..023f816a695 100644
--- a/examples/geo3k_vlm/run_geo3k_vlm_sft.sh
+++ b/examples/geo3k_vlm/run_geo3k_vlm_sft.sh
@@ -152,7 +152,8 @@ else
    MILES_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/../.." &>/dev/null && pwd)"
    MODEL_ARGS_FILE=$(echo "$MODEL_NAME" | sed 's/-Instruct//g; s/-Thinking//g; s/Qwen3-VL-/qwen3-/g; s/-2B/-1.7B/g')
    # VL models require rotary-base 5000000
-   MODEL_ARGS_ROTARY_BASE=5000000 source "${MILES_DIR}/scripts/models/${MODEL_ARGS_FILE}.sh"
+   MODEL_ARGS_LINE="$(MODEL_ARGS_ROTARY_BASE=5000000 python3 "${MILES_DIR}/miles/utils/external_utils/model_args_utils.py" "${MODEL_ARGS_FILE}")" || exit 1
+   read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 fi
 
 # Start Ray if not using external Ray
diff --git a/examples/infra_features/low_precision/run-kimi-k2-Thinking-int4.sh b/examples/infra_features/low_precision/run-kimi-k2-Thinking-int4.sh
index f7fa4840e2f..0572681da73 100644
--- a/examples/infra_features/low_precision/run-kimi-k2-Thinking-int4.sh
+++ b/examples/infra_features/low_precision/run-kimi-k2-Thinking-int4.sh
@@ -24,8 +24,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/../../../scripts/models/kimi-k2-thinking.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../../../miles/utils/external_utils/model_args_utils.py" "kimi-k2-thinking")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint /root/Kimi-K2-Thinking/
    --ref-load /root/Kimi-K2_thinking_torch_dist/
diff --git a/examples/infra_features/low_precision/run-moonlight-16B-A3B-int4.sh b/examples/infra_features/low_precision/run-moonlight-16B-A3B-int4.sh
index 17b81f2ee83..f15a4245710 100644
--- a/examples/infra_features/low_precision/run-moonlight-16B-A3B-int4.sh
+++ b/examples/infra_features/low_precision/run-moonlight-16B-A3B-int4.sh
@@ -25,8 +25,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/../../../scripts/models/moonlight.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../../../miles/utils/external_utils/model_args_utils.py" "moonlight")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint /root/Moonlight-16B-A3B-Instruct-INT4
    --ref-load /root/Moonlight-16B-A3B-Instruct-INT4_torch_dist
diff --git a/examples/infra_features/low_precision/run-qwen3-235B-A22B-int4.sh b/examples/infra_features/low_precision/run-qwen3-235B-A22B-int4.sh
index 490f8e41418..54a47fba575 100644
--- a/examples/infra_features/low_precision/run-qwen3-235B-A22B-int4.sh
+++ b/examples/infra_features/low_precision/run-qwen3-235B-A22B-int4.sh
@@ -24,8 +24,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/../../../scripts/models/qwen3-235B-A22B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../../../miles/utils/external_utils/model_args_utils.py" "qwen3-235B-A22B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint /root/Qwen3-235B-A22B-INT4/
    --ref-load /root/Qwen3-235B-A22B_torch_dist/
diff --git a/examples/infra_features/low_precision/run-qwen3-30B-A3B-int4.sh b/examples/infra_features/low_precision/run-qwen3-30B-A3B-int4.sh
index 0ff20072ec8..9b633d938fc 100644
--- a/examples/infra_features/low_precision/run-qwen3-30B-A3B-int4.sh
+++ b/examples/infra_features/low_precision/run-qwen3-30B-A3B-int4.sh
@@ -24,8 +24,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/../../../scripts/models/qwen3-30B-A3B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../../../miles/utils/external_utils/model_args_utils.py" "qwen3-30B-A3B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint /root/Qwen3-30B-A3B-INT4/
    --ref-load /root/Qwen3-30B-A3B_torch_dist/
diff --git a/examples/infra_features/low_precision/run-qwen3-30b-a3b-fp8-two-nodes.sh b/examples/infra_features/low_precision/run-qwen3-30b-a3b-fp8-two-nodes.sh
index 0f6fbf6b5be..19cca96e043 100644
--- a/examples/infra_features/low_precision/run-qwen3-30b-a3b-fp8-two-nodes.sh
+++ b/examples/infra_features/low_precision/run-qwen3-30b-a3b-fp8-two-nodes.sh
@@ -25,7 +25,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/../../../scripts/models/qwen3-30B-A3B.sh"
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../../../miles/utils/external_utils/model_args_utils.py" "qwen3-30B-A3B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 
 # Base directory for checkpoints and related files (adjust if necessary)
 BASE_DIR="/root" 
diff --git a/examples/infra_features/low_precision/run-qwen3-4b-fp8.sh b/examples/infra_features/low_precision/run-qwen3-4b-fp8.sh
index bf8f6407aeb..205cda89a72 100644
--- a/examples/infra_features/low_precision/run-qwen3-4b-fp8.sh
+++ b/examples/infra_features/low_precision/run-qwen3-4b-fp8.sh
@@ -24,7 +24,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/../../../scripts/models/qwen3-4B.sh"
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../../../miles/utils/external_utils/model_args_utils.py" "qwen3-4B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 
 CKPT_ARGS=(
    --hf-checkpoint /root/Qwen3-4B-FP8
diff --git a/examples/infra_features/p2p_weight_transfer/run-glm4.5-air-8node-profile.sh b/examples/infra_features/p2p_weight_transfer/run-glm4.5-air-8node-profile.sh
index b6981ced2f3..652b310a393 100755
--- a/examples/infra_features/p2p_weight_transfer/run-glm4.5-air-8node-profile.sh
+++ b/examples/infra_features/p2p_weight_transfer/run-glm4.5-air-8node-profile.sh
@@ -72,8 +72,8 @@ MODEL_NAME="GLM-4.5-Air"
 MODEL_TYPE="glm4.5-106B-A12B"
 
 MILES_ROOT="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/../../.." &>/dev/null && pwd)"
-source "${MILES_ROOT}/scripts/models/${MODEL_TYPE}.sh"
-
+MODEL_ARGS_LINE="$(python3 "${MILES_ROOT}/miles/utils/external_utils/model_args_utils.py" "${MODEL_TYPE}")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 # Rotary base override
 export MODEL_ARGS_ROTARY_BASE=1000000
 
diff --git a/examples/infra_features/p2p_weight_transfer/run-glm4.7-flash-2node-profile.sh b/examples/infra_features/p2p_weight_transfer/run-glm4.7-flash-2node-profile.sh
index b2604b5c24f..9062fa0cfd3 100644
--- a/examples/infra_features/p2p_weight_transfer/run-glm4.7-flash-2node-profile.sh
+++ b/examples/infra_features/p2p_weight_transfer/run-glm4.7-flash-2node-profile.sh
@@ -80,8 +80,8 @@ MODEL_TYPE="glm4.7-flash"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
 MILES_ROOT="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/../../.." &>/dev/null && pwd)"
-source "${MILES_ROOT}/scripts/models/${MODEL_TYPE}.sh"
-
+MODEL_ARGS_LINE="$(python3 "${MILES_ROOT}/miles/utils/external_utils/model_args_utils.py" "${MODEL_TYPE}")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 # ---------------------------------------------------------------------------
 # Determine modes to run
 # ---------------------------------------------------------------------------
diff --git a/examples/infra_features/p2p_weight_transfer/run-glm5-disagg-profile.sh b/examples/infra_features/p2p_weight_transfer/run-glm5-disagg-profile.sh
index 0a279c28dc0..7ce1df610f8 100755
--- a/examples/infra_features/p2p_weight_transfer/run-glm5-disagg-profile.sh
+++ b/examples/infra_features/p2p_weight_transfer/run-glm5-disagg-profile.sh
@@ -116,8 +116,8 @@ esac
 NUM_TRAIN_NODES=$((NUM_TRAIN_GPUS / GPUS_PER_NODE))
 
 MILES_ROOT="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/../../.." &>/dev/null && pwd)"
-source "${MILES_ROOT}/scripts/models/${MODEL_TYPE}.sh"
-
+MODEL_ARGS_LINE="$(python3 "${MILES_ROOT}/miles/utils/external_utils/model_args_utils.py" "${MODEL_TYPE}")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 echo ""
 echo "============================================================"
 echo "  Model      : ${MODEL_NAME} (${MODEL_TYPE})"
diff --git a/examples/infra_features/p2p_weight_transfer/run-kimi-k2-64node-profile.sh b/examples/infra_features/p2p_weight_transfer/run-kimi-k2-64node-profile.sh
index 8dc4e008533..ddcdd9c4473 100644
--- a/examples/infra_features/p2p_weight_transfer/run-kimi-k2-64node-profile.sh
+++ b/examples/infra_features/p2p_weight_transfer/run-kimi-k2-64node-profile.sh
@@ -85,9 +85,8 @@ MODEL_TYPE="kimi-k2"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
 MILES_ROOT="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/../../.." &>/dev/null && pwd)"
-source "${MILES_ROOT}/scripts/models/${MODEL_TYPE}.sh"
-
-
+MODEL_ARGS_LINE="$(python3 "${MILES_ROOT}/miles/utils/external_utils/model_args_utils.py" "${MODEL_TYPE}")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 # ---------------------------------------------------------------------------
 # Determine modes to run
 # ---------------------------------------------------------------------------
diff --git a/examples/infra_features/p2p_weight_transfer/run-qwen3-235B-A22B-16node-profile.sh b/examples/infra_features/p2p_weight_transfer/run-qwen3-235B-A22B-16node-profile.sh
index 129d0abb990..57a28d1ca7c 100644
--- a/examples/infra_features/p2p_weight_transfer/run-qwen3-235B-A22B-16node-profile.sh
+++ b/examples/infra_features/p2p_weight_transfer/run-qwen3-235B-A22B-16node-profile.sh
@@ -82,9 +82,8 @@ MODEL_TYPE="qwen3-235B-A22B"
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
 MILES_ROOT="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/../../.." &>/dev/null && pwd)"
 export MODEL_ARGS_ROTARY_BASE=5000000
-source "${MILES_ROOT}/scripts/models/${MODEL_TYPE}.sh"
-
-
+MODEL_ARGS_LINE="$(python3 "${MILES_ROOT}/miles/utils/external_utils/model_args_utils.py" "${MODEL_TYPE}")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 # ---------------------------------------------------------------------------
 # Determine modes to run
 # ---------------------------------------------------------------------------
diff --git a/examples/infra_features/p2p_weight_transfer/run-qwen3-30B-A3B-4node-profile.sh b/examples/infra_features/p2p_weight_transfer/run-qwen3-30B-A3B-4node-profile.sh
index e17a48445fd..43b610f3f6e 100644
--- a/examples/infra_features/p2p_weight_transfer/run-qwen3-30B-A3B-4node-profile.sh
+++ b/examples/infra_features/p2p_weight_transfer/run-qwen3-30B-A3B-4node-profile.sh
@@ -84,8 +84,8 @@ MODEL_TYPE="qwen3-30B-A3B"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
 MILES_ROOT="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/../../.." &>/dev/null && pwd)"
-source "${MILES_ROOT}/scripts/models/${MODEL_TYPE}.sh"
-
+MODEL_ARGS_LINE="$(python3 "${MILES_ROOT}/miles/utils/external_utils/model_args_utils.py" "${MODEL_TYPE}")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 # Rotary base override
 export MODEL_ARGS_ROTARY_BASE=1000000
 
diff --git a/examples/infra_features/p2p_weight_transfer/run.py b/examples/infra_features/p2p_weight_transfer/run.py
index dd14e9b145f..1014b13a97b 100644
--- a/examples/infra_features/p2p_weight_transfer/run.py
+++ b/examples/infra_features/p2p_weight_transfer/run.py
@@ -35,7 +35,7 @@ class PrepareConfig:
     """Configuration for the `prepare` subcommand."""
 
     hf_repo: str
-    model_type: str  # megatron model type (maps to scripts/models/<type>.sh)
+    model_type: str  # megatron model type (maps to scripts/models/<type>.py)
     datasets: list[str] = field(default_factory=lambda: ["zhuzilin/dapo-math-17k"])
     convert_gpus_per_node: int = 8
     convert_multinode: bool = False
@@ -1101,7 +1101,11 @@ def cmd_run(
 def build_model_args_command(cfg: RunConfig) -> str:
     """A shell snippet leaving MODEL_ARGS set; the knobs must reach it, not only ray's runtime env."""
     prefix = "".join(f"{name}={shlex.quote(value)} " for name, value in build_model_args_env(cfg).items())
-    return f'{prefix}source "{MILES_ROOT}/scripts/models/{cfg.model_type}.sh"'
+    return (
+        f'MODEL_ARGS_LINE="$({prefix}python3 "{MILES_ROOT}/miles/utils/external_utils/model_args_utils.py"'
+        f' {cfg.model_type})" || exit 1; '
+        'read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"'
+    )
 
 
 def build_model_args_env(cfg: RunConfig) -> dict[str, str]:
diff --git a/examples/infra_features/train_infer_mismatch_helper/run-qwen3-4b-mis.sh b/examples/infra_features/train_infer_mismatch_helper/run-qwen3-4b-mis.sh
index 4a3f0aeda77..1390d3ac836 100644
--- a/examples/infra_features/train_infer_mismatch_helper/run-qwen3-4b-mis.sh
+++ b/examples/infra_features/train_infer_mismatch_helper/run-qwen3-4b-mis.sh
@@ -24,7 +24,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/../../../scripts/models/qwen3-4B.sh"
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../../../miles/utils/external_utils/model_args_utils.py" "qwen3-4B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 
 CKPT_ARGS=(
    --hf-checkpoint /root/Qwen3-4B
diff --git a/examples/lora/dev.sh b/examples/lora/dev.sh
index 80648cfb810..0b2145efb94 100644
--- a/examples/lora/dev.sh
+++ b/examples/lora/dev.sh
@@ -19,8 +19,8 @@ set -ex
 
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/../../scripts/models/qwen2.5-3B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../../miles/utils/external_utils/model_args_utils.py" "qwen2.5-3B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint /root/Qwen2.5-3B-Instruct/
    --megatron-to-hf-mode bridge
diff --git a/examples/lora/run-gpt-oss-20B-megatron-moe-lora.sh b/examples/lora/run-gpt-oss-20B-megatron-moe-lora.sh
index e26d663fa23..098a0e76865 100644
--- a/examples/lora/run-gpt-oss-20B-megatron-moe-lora.sh
+++ b/examples/lora/run-gpt-oss-20B-megatron-moe-lora.sh
@@ -16,8 +16,8 @@ GPUS_PER_NODE=$(echo "$CUDA_VISIBLE_DEVICES" | tr ',' '\n' | wc -l)
 
 # Load model architecture config
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/../../scripts/models/gpt-oss-20b.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../../miles/utils/external_utils/model_args_utils.py" "gpt-oss-20b")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint /root/models/gpt-oss-20b
    --megatron-to-hf-mode bridge
diff --git a/examples/lora/run-kimi-k25-megatron-lora.sh b/examples/lora/run-kimi-k25-megatron-lora.sh
index f43f141dabd..ec25afdf91c 100755
--- a/examples/lora/run-kimi-k25-megatron-lora.sh
+++ b/examples/lora/run-kimi-k25-megatron-lora.sh
@@ -28,8 +28,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/../../scripts/models/kimi-k2-thinking.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../../miles/utils/external_utils/model_args_utils.py" "kimi-k2-thinking")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint $BASE_DIR/Kimi-K2.5-int4
    --ref-load $BASE_DIR/Kimi-K2.5-bf16
diff --git a/examples/lora/run-qwen2.5-0.5B-megatron-lora.sh b/examples/lora/run-qwen2.5-0.5B-megatron-lora.sh
index bc287acd695..be29ca03790 100644
--- a/examples/lora/run-qwen2.5-0.5B-megatron-lora.sh
+++ b/examples/lora/run-qwen2.5-0.5B-megatron-lora.sh
@@ -18,8 +18,8 @@ set -ex
 
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/../../scripts/models/qwen2.5-0.5B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../../miles/utils/external_utils/model_args_utils.py" "qwen2.5-0.5B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint /root/Qwen2.5-0.5B-Instruct/
    --megatron-to-hf-mode bridge
diff --git a/examples/lora/run-qwen2.5-3B-megatron-lora-disaggregated-multi-node.sh b/examples/lora/run-qwen2.5-3B-megatron-lora-disaggregated-multi-node.sh
index 592c77148a2..ab8556f719d 100644
--- a/examples/lora/run-qwen2.5-3B-megatron-lora-disaggregated-multi-node.sh
+++ b/examples/lora/run-qwen2.5-3B-megatron-lora-disaggregated-multi-node.sh
@@ -134,8 +134,8 @@ pkill -9 python || true
 set -ex
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/../../scripts/models/qwen2.5-3B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../../miles/utils/external_utils/model_args_utils.py" "qwen2.5-3B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint /root/Qwen2.5-3B-Instruct/
    --megatron-to-hf-mode bridge
diff --git a/examples/lora/run-qwen2.5-3B-megatron-lora-disaggregated.sh b/examples/lora/run-qwen2.5-3B-megatron-lora-disaggregated.sh
index f04fabf65bf..7687a83d3c2 100644
--- a/examples/lora/run-qwen2.5-3B-megatron-lora-disaggregated.sh
+++ b/examples/lora/run-qwen2.5-3B-megatron-lora-disaggregated.sh
@@ -18,8 +18,8 @@ set -ex
 
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/../../scripts/models/qwen2.5-3B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../../miles/utils/external_utils/model_args_utils.py" "qwen2.5-3B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint /root/Qwen2.5-3B-Instruct/
    --megatron-to-hf-mode bridge
diff --git a/examples/lora/run-qwen3-4B-megatron-lora.sh b/examples/lora/run-qwen3-4B-megatron-lora.sh
index 48ed58344e6..e47a50fc1f5 100644
--- a/examples/lora/run-qwen3-4B-megatron-lora.sh
+++ b/examples/lora/run-qwen3-4B-megatron-lora.sh
@@ -30,7 +30,8 @@ echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
 REPO_ROOT="$(cd -- "${SCRIPT_DIR}/../.." &>/dev/null && pwd)"
-source "${REPO_ROOT}/scripts/models/qwen3-4B.sh"
+MODEL_ARGS_LINE="$(python3 "${REPO_ROOT}/miles/utils/external_utils/model_args_utils.py" "qwen3-4B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 
 # Store eval/delegate settings in a YAML config similar to examples/experimental/eval_multi_task.
 # EVAL_CONFIG_PATH=${SKILLS_EVAL_CONFIG_PATH:-"${REPO_ROOT}/examples/experimental/eval/scripts/multi_tasks.yaml"}
diff --git a/examples/lora/run-qwen3-4b-megatron-lora-result.sh b/examples/lora/run-qwen3-4b-megatron-lora-result.sh
index 0d76d807cb3..624a76892d5 100644
--- a/examples/lora/run-qwen3-4b-megatron-lora-result.sh
+++ b/examples/lora/run-qwen3-4b-megatron-lora-result.sh
@@ -33,8 +33,8 @@ echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 LR=2e-5
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/../../scripts/models/qwen3-4B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../../miles/utils/external_utils/model_args_utils.py" "qwen3-4B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint /root/Qwen3-4B
    --save /root/Qwen3-4B-lora-ckpt
diff --git a/examples/on_policy_distillation/run-qwen3-8B-opd-megatron.sh b/examples/on_policy_distillation/run-qwen3-8B-opd-megatron.sh
index 0e971aae269..df65334a7d2 100644
--- a/examples/on_policy_distillation/run-qwen3-8B-opd-megatron.sh
+++ b/examples/on_policy_distillation/run-qwen3-8B-opd-megatron.sh
@@ -22,9 +22,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/../../scripts/models/qwen3-8B.sh"
-
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../../miles/utils/external_utils/model_args_utils.py" "qwen3-8B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint /root/Qwen3-8B
    --ref-load /root/Qwen3-8B_torch_dist
diff --git a/examples/on_policy_distillation/run-qwen3-8B-opd-multi-teacher.sh b/examples/on_policy_distillation/run-qwen3-8B-opd-multi-teacher.sh
index fad2c0aadee..1dae8db8a02 100644
--- a/examples/on_policy_distillation/run-qwen3-8B-opd-multi-teacher.sh
+++ b/examples/on_policy_distillation/run-qwen3-8B-opd-multi-teacher.sh
@@ -97,7 +97,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/../../scripts/models/qwen3-8B.sh"
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../../miles/utils/external_utils/model_args_utils.py" "qwen3-8B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 
 
 CKPT_ARGS=(
diff --git a/examples/on_policy_distillation/run-qwen3-8B-opd.sh b/examples/on_policy_distillation/run-qwen3-8B-opd.sh
index 1389138133f..953c380c00d 100644
--- a/examples/on_policy_distillation/run-qwen3-8B-opd.sh
+++ b/examples/on_policy_distillation/run-qwen3-8B-opd.sh
@@ -45,9 +45,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/../../scripts/models/qwen3-8B.sh"
-
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../../miles/utils/external_utils/model_args_utils.py" "qwen3-8B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint /root/Qwen3-8B
    --ref-load /root/Qwen3-8B_torch_dist
diff --git a/miles/utils/debug_utils/run_megatron/cli/commands/args.py b/miles/utils/debug_utils/run_megatron/cli/commands/args.py
index a851b0e400c..818a709eaaf 100644
--- a/miles/utils/debug_utils/run_megatron/cli/commands/args.py
+++ b/miles/utils/debug_utils/run_megatron/cli/commands/args.py
@@ -15,7 +15,7 @@ def _field(
 
 @dataclasses.dataclass
 class CommonRunArgs:
-    model_type: str = _field(help="Model type matching scripts/models/{model_type}.sh")
+    model_type: str = _field(help="Model type matching scripts/models/{model_type}.py")
     hf_checkpoint: Path = _field(help="HuggingFace checkpoint path")
     ref_load: Path | None = _field(default=None, help="Megatron checkpoint path")
     sp: bool = _field(default=False, help="Enable sequence parallelism")
diff --git a/miles/utils/debug_utils/run_megatron/cli/commands/run.py b/miles/utils/debug_utils/run_megatron/cli/commands/run.py
index 4f1a53659f8..e83e98e1499 100644
--- a/miles/utils/debug_utils/run_megatron/cli/commands/run.py
+++ b/miles/utils/debug_utils/run_megatron/cli/commands/run.py
@@ -7,7 +7,7 @@
 
 from miles.utils.debug_utils.run_megatron.cli.commands.args import RunArgs
 from miles.utils.debug_utils.run_megatron.cli.parallel_utils import ParallelConfig
-from miles.utils.debug_utils.run_megatron.cli.path_utils import resolve_megatron_path, resolve_model_script
+from miles.utils.debug_utils.run_megatron.cli.path_utils import resolve_megatron_path
 from miles.utils.debug_utils.run_megatron.cli.prompt_utils import (
     PromptConfig,
     generate_token_ids,
@@ -19,7 +19,8 @@
     build_worker_args,
 )
 from miles.utils.debug_utils.run_megatron.worker.script_args import WorkerScriptArgs
-from miles.utils.external_utils.exec_command import exec_command_cpu, exec_command_gpu
+from miles.utils.external_utils.exec_command import exec_command_gpu
+from miles.utils.external_utils.model_args_utils import load_model_args
 from miles.utils.typer_utils import dataclass_cli
 
 
@@ -94,12 +95,7 @@ def run(args: RunArgs) -> None:
 
 
 def show_model_args(
-    model_type: Annotated[str, typer.Option(help="Model type matching scripts/models/{model_type}.sh")],
+    model_type: Annotated[str, typer.Option(help="Model type matching scripts/models/{model_type}.py")],
 ) -> None:
     """Show the MODEL_ARGS for a given model type (debug helper)."""
-    output: str | None = exec_command_cpu(
-        f'source "{resolve_model_script(model_type)}" && echo "${{MODEL_ARGS[@]}}"',
-        capture_output=True,
-    )
-    if output:
-        print(output.strip())
+    print(load_model_args(model_type))
diff --git a/miles/utils/debug_utils/run_megatron/cli/path_utils.py b/miles/utils/debug_utils/run_megatron/cli/path_utils.py
index b34f91d1459..810f0539632 100644
--- a/miles/utils/debug_utils/run_megatron/cli/path_utils.py
+++ b/miles/utils/debug_utils/run_megatron/cli/path_utils.py
@@ -19,7 +19,7 @@ def resolve_megatron_path(megatron_path: Path | None) -> Path:
 
 def resolve_model_script(model_type: str) -> Path:
     repo_base: Path = _resolve_repo_base()
-    script: Path = repo_base / "scripts" / "models" / f"{model_type}.sh"
+    script: Path = repo_base / "scripts" / "models" / f"{model_type}.py"
     if not script.exists():
         raise typer.BadParameter(f"Model script not found: {script}")
     return script
diff --git a/miles/utils/debug_utils/run_megatron/cli/worker_executor.py b/miles/utils/debug_utils/run_megatron/cli/worker_executor.py
index 211ae089067..5bb6644bbf9 100644
--- a/miles/utils/debug_utils/run_megatron/cli/worker_executor.py
+++ b/miles/utils/debug_utils/run_megatron/cli/worker_executor.py
@@ -3,8 +3,8 @@
 from pathlib import Path
 
 from miles.utils.debug_utils.run_megatron.cli.parallel_utils import ParallelConfig
-from miles.utils.debug_utils.run_megatron.cli.path_utils import resolve_model_script
 from miles.utils.debug_utils.run_megatron.worker.script_args import WORKER_SCRIPT_ARGS_BRIDGE, WorkerScriptArgs
+from miles.utils.external_utils.model_args_utils import load_model_args
 
 
 def build_torchrun_cmd(
@@ -15,16 +15,14 @@ def build_torchrun_cmd(
     worker_args: str,
 ) -> str:
     """Build the full shell command to launch the worker via torchrun."""
-    model_script: Path = resolve_model_script(model_type)
     worker_module: str = "miles.utils.debug_utils.run_megatron.worker.main"
 
     cmd: str = (
-        f'source "{model_script}" && '
         f"PYTHONPATH={megatron_path}:$PYTHONPATH "
         f"CUDA_DEVICE_MAX_CONNECTIONS=1 "
         f"torchrun --nproc-per-node {nproc} "
         f"-m {worker_module} "
-        f"${{MODEL_ARGS[@]}} "
+        f"{load_model_args(model_type)} "
         f"--hidden-dropout 0 --attention-dropout 0 "
         f"{worker_args}"
     )
diff --git a/miles/utils/external_utils/command_utils.py b/miles/utils/external_utils/command_utils.py
index d296734ba69..ae0e96ee2a6 100644
--- a/miles/utils/external_utils/command_utils.py
+++ b/miles/utils/external_utils/command_utils.py
@@ -9,12 +9,12 @@
 import random
 import shlex
 import socket
-import subprocess
 from dataclasses import dataclass, field
 from functools import partial
 from pathlib import Path
 
 from miles.utils.external_utils.exec_command import exec_command_cpu, exec_command_gpu, exec_command_multi_node
+from miles.utils.external_utils.model_args_utils import load_model_args
 from miles.utils.file_arg_utils import PSEUDO_FILE_PREFIX
 from miles.utils.http_utils import wait_for_server_ready
 from miles.utils.typer_utils import dataclass_cli
@@ -32,18 +32,6 @@ def _pythonpath_with_sources(megatron_path: str, *additional_pythonpaths: str |
     return os.pathsep.join(dict.fromkeys(entries))
 
 
-def load_model_args(megatron_model_type: str) -> list[str]:
-    """Expand the MODEL_ARGS array that scripts/models/<megatron_model_type>.sh declares."""
-    script = f"{repo_base_dir}/scripts/models/{megatron_model_type}.sh"
-    assert os.path.exists(script), f"no model args script at {script}"
-    expansion = f'source {shlex.quote(script)} && printf "%s\\0" "${{MODEL_ARGS[@]}}"'
-    result = subprocess.run(["bash", "-c", expansion], capture_output=True, text=True, check=True)
-    tokens = result.stdout.split("\0")[:-1]
-    for token in tokens:
-        assert token.split() == [token], f"model args token must be one whitespace-free word: {token!r}"
-    return tokens
-
-
 def convert_checkpoint(
     model_name,
     megatron_model_type,
@@ -81,7 +69,7 @@ def convert_checkpoint(
         f"--nproc-per-node {num_gpus_per_node} "
         f"{multinode_args}"
         f"{repo_base_dir}/tools/convert_hf_to_torch_dist.py "
-        f"{' '.join(load_model_args(megatron_model_type))} "
+        f"{load_model_args(megatron_model_type)} "
         f"--hf-checkpoint {hf_checkpoint} "
         f"--save {path_dst} "
         f"{extra_args}"
@@ -207,7 +195,7 @@ def execute_train(
     runtime_env_json = json.dumps({"env_vars": runtime_env_vars})
 
     if get_bool_env_var("MILES_SCRIPT_ENABLE_RAY_SUBMIT", "1"):
-        model_args = " ".join(load_model_args(megatron_model_type)) if megatron_model_type is not None else ""
+        model_args = load_model_args(megatron_model_type) if megatron_model_type is not None else ""
         exec_command_cpu(
             f"export no_proxy=127.0.0.1 && export PYTHONUNBUFFERED=1 && "
             f"""ray job submit {'' if 'RAY_ADDRESS' in os.environ else '--address="http://127.0.0.1:8265" '}"""
diff --git a/miles/utils/external_utils/model_args_utils.py b/miles/utils/external_utils/model_args_utils.py
new file mode 100644
index 00000000000..698f38aa2ad
--- /dev/null
+++ b/miles/utils/external_utils/model_args_utils.py
@@ -0,0 +1,59 @@
+import importlib.util
+import sys
+from pathlib import Path
+from types import ModuleType
+
+REPO_ROOT = Path(__file__).resolve().parents[3]
+MODEL_SCRIPT_DIR = REPO_ROOT / "scripts" / "models"
+
+
+# ==================== loading a model script ====================
+
+
+def load_model_args(model_type: str, model_script_dir: Path | None = None, **kwargs: object) -> str:
+    """Collapse scripts/models/<model_type>.py to one line; a newline would truncate the shell's read -ra."""
+    path = (model_script_dir or MODEL_SCRIPT_DIR) / f"{model_type}.py"
+    assert path.exists(), f"no model args script at {path}"
+    sys.modules.setdefault("model_args_utils", sys.modules[__name__])
+    module = import_module_from_path(path, f"miles_model_args_{path.stem.replace('.', '_').replace('-', '_')}")
+    args = " ".join(module.model_args(**kwargs).split())
+    assert args, f"{path} declared no model args"
+    return args
+
+
+def load_sibling_model_args(model_script: str, model_type: str, **kwargs: object) -> str:
+    """Load the model a variant is derived from, out of the same checkout as the variant itself."""
+    return load_model_args(model_type, model_script_dir=Path(model_script).resolve().parent, **kwargs)
+
+
+# ==================== what a model script may call ====================
+
+
+def moe_layer_freq(*, nlayers: int, first_k_dense_replace: int) -> str:
+    """Render megatron's --moe-layer-freq pattern: the first K layers dense, the rest MoE."""
+    dense = min(first_k_dense_replace, nlayers)
+    return "[" + ",".join(["0"] * dense + ["1"] * (nlayers - dense)) + "]"
+
+
+# ==================== importing a file by path ====================
+
+
+def import_module_from_path(path: Path, module_name: str) -> ModuleType:
+    """Import a python file that is not reachable as a dotted module path."""
+    spec = importlib.util.spec_from_file_location(module_name, path)
+    assert spec is not None and spec.loader is not None, f"cannot load {path}"
+    module = importlib.util.module_from_spec(spec)
+    sys.modules[module_name] = module
+    try:
+        spec.loader.exec_module(module)
+    finally:
+        del sys.modules[module_name]
+    return module
+
+
+# ==================== command line ====================
+
+
+if __name__ == "__main__":
+    (_MODEL_TYPE,) = sys.argv[1:]
+    print(load_model_args(_MODEL_TYPE))
diff --git a/scripts/amd/run-qwen3-4B-amd.sh b/scripts/amd/run-qwen3-4B-amd.sh
index d3251fcf50a..9bb31865373 100644
--- a/scripts/amd/run-qwen3-4B-amd.sh
+++ b/scripts/amd/run-qwen3-4B-amd.sh
@@ -30,8 +30,8 @@ if [[ -n "${CUDA_VISIBLE_DEVICES:-}" ]]; then
 fi
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/.." &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/qwen3-4B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "qwen3-4B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint /root/Qwen3-4B
    --ref-load /root/Qwen3-4B_torch_dist
diff --git a/scripts/models/deepseek-v3-20layer.py b/scripts/models/deepseek-v3-20layer.py
new file mode 100644
index 00000000000..19bc7a48c94
--- /dev/null
+++ b/scripts/models/deepseek-v3-20layer.py
@@ -0,0 +1,5 @@
+from model_args_utils import load_sibling_model_args
+
+
+def model_args() -> str:
+    return load_sibling_model_args(__file__, "deepseek-v3", nlayers=20)
diff --git a/scripts/models/deepseek-v3-20layer.sh b/scripts/models/deepseek-v3-20layer.sh
deleted file mode 100644
index 6fdde1820c8..00000000000
--- a/scripts/models/deepseek-v3-20layer.sh
+++ /dev/null
@@ -1 +0,0 @@
-MODEL_ARGS_NUM_LAYERS=20 source "$(dirname -- "${BASH_SOURCE[0]}")/deepseek-v3.sh"
diff --git a/scripts/models/deepseek-v3-5layer.py b/scripts/models/deepseek-v3-5layer.py
new file mode 100644
index 00000000000..910b1b951f8
--- /dev/null
+++ b/scripts/models/deepseek-v3-5layer.py
@@ -0,0 +1,5 @@
+from model_args_utils import load_sibling_model_args
+
+
+def model_args() -> str:
+    return load_sibling_model_args(__file__, "deepseek-v3", nlayers=5)
diff --git a/scripts/models/deepseek-v3-5layer.sh b/scripts/models/deepseek-v3-5layer.sh
deleted file mode 100644
index a5e5d2522ce..00000000000
--- a/scripts/models/deepseek-v3-5layer.sh
+++ /dev/null
@@ -1 +0,0 @@
-MODEL_ARGS_NUM_LAYERS=5 source "$(dirname -- "${BASH_SOURCE[0]}")/deepseek-v3.sh"
diff --git a/scripts/models/deepseek-v3.py b/scripts/models/deepseek-v3.py
new file mode 100644
index 00000000000..ce0bc96f2ef
--- /dev/null
+++ b/scripts/models/deepseek-v3.py
@@ -0,0 +1,56 @@
+import os
+
+from model_args_utils import moe_layer_freq
+
+
+FIRST_K_DENSE_REPLACE = 3
+
+
+def model_args(nlayers: int | None = None) -> str:
+    nlayers = nlayers if nlayers is not None else int(os.environ.get("MODEL_ARGS_NUM_LAYERS") or 61)
+    return (
+        "--disable-bias-linear "
+        f"--num-layers {nlayers} "
+        "--hidden-size 7168 "
+        "--ffn-hidden-size 18432 "
+        "--num-attention-heads 128 "
+        "--kv-channels 128 "
+        "--normalization RMSNorm "
+        "--position-embedding-type rope "
+        "--norm-epsilon 1e-6 "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--vocab-size 129280 "
+        "--multi-latent-attention "
+        "--q-lora-rank 1536 "
+        "--kv-lora-rank 512 "
+        "--qk-head-dim 128 "
+        "--qk-pos-emb-head-dim 64 "
+        "--v-head-dim 128 "
+        "--qk-layernorm "
+        "--rotary-scaling-factor 40 "
+        "--rotary-base 10000 "
+        "--mscale 1.0 "
+        "--mscale-all-dim 1.0 "
+        "--attention-softmax-in-fp32 "
+        "--no-rope-fusion "
+        # moe
+        "--num-experts 256 "
+        f"--moe-layer-freq {moe_layer_freq(nlayers=nlayers, first_k_dense_replace=FIRST_K_DENSE_REPLACE)} "
+        "--moe-ffn-hidden-size 2048 "
+        "--moe-router-topk 8 "
+        "--moe-shared-expert-intermediate-size 2048 "
+        "--moe-router-pre-softmax "
+        "--moe-router-score-function sigmoid "
+        "--moe-router-enable-expert-bias "
+        "--moe-router-load-balancing-type seq_aux_loss "
+        "--moe-token-dispatcher-type alltoall "
+        "--moe-aux-loss-coeff 0 "
+        "--moe-router-bias-update-rate 0 "
+        "--moe-router-group-topk 4 "
+        "--moe-router-num-groups 8 "
+        "--moe-grouped-gemm "
+        "--moe-router-topk-scaling-factor 2.5 "
+        "--moe-router-dtype fp32 "
+        "--moe-permute-fusion "
+    )
diff --git a/scripts/models/deepseek-v3.sh b/scripts/models/deepseek-v3.sh
deleted file mode 100644
index 8c50d2c9405..00000000000
--- a/scripts/models/deepseek-v3.sh
+++ /dev/null
@@ -1,63 +0,0 @@
-NLAYERS="${MODEL_ARGS_NUM_LAYERS:-61}"
-FIRST_K_DENSE_REPLACE=3
-
-arr=()
-for ((i=0; i<NLAYERS; i++)); do
-  if (( i < FIRST_K_DENSE_REPLACE )); then
-    arr+=(0)
-  else
-    arr+=(1)
-  fi
-done
-
-printf -v MOE_LAYER_FREQ "[%s]" "$(IFS=', '; echo "${arr[*]}")"
-
-# deepseek r1
-MODEL_ARGS=(
-    --disable-bias-linear
-    --num-layers $NLAYERS
-    --hidden-size 7168
-    --ffn-hidden-size 18432
-    --num-attention-heads 128
-    --kv-channels 128
-    --normalization RMSNorm
-    --position-embedding-type rope
-    --norm-epsilon 1e-6
-    --swiglu
-    --untie-embeddings-and-output-weights
-    --vocab-size 129280
-
-    --multi-latent-attention
-    --q-lora-rank 1536
-    --kv-lora-rank 512
-    --qk-head-dim 128
-    --qk-pos-emb-head-dim 64
-    --v-head-dim 128
-    --qk-layernorm
-    --rotary-scaling-factor 40
-    --rotary-base 10000
-    --mscale 1.0
-    --mscale-all-dim 1.0
-    --attention-softmax-in-fp32
-    --no-rope-fusion
-
-    # moe
-    --num-experts 256
-    --moe-layer-freq "$MOE_LAYER_FREQ"
-    --moe-ffn-hidden-size 2048
-    --moe-router-topk 8
-    --moe-shared-expert-intermediate-size 2048
-    --moe-router-pre-softmax
-    --moe-router-score-function sigmoid
-    --moe-router-enable-expert-bias
-    --moe-router-load-balancing-type seq_aux_loss
-    --moe-token-dispatcher-type alltoall
-    --moe-aux-loss-coeff 0
-    --moe-router-bias-update-rate 0
-    --moe-router-group-topk 4
-    --moe-router-num-groups 8
-    --moe-grouped-gemm
-    --moe-router-topk-scaling-factor 2.5
-    --moe-router-dtype fp32
-    --moe-permute-fusion
-)
\ No newline at end of file
diff --git a/scripts/models/deepseek-v32-5layer.py b/scripts/models/deepseek-v32-5layer.py
new file mode 100644
index 00000000000..382e8ce357f
--- /dev/null
+++ b/scripts/models/deepseek-v32-5layer.py
@@ -0,0 +1,5 @@
+from model_args_utils import load_sibling_model_args
+
+
+def model_args() -> str:
+    return load_sibling_model_args(__file__, "deepseek-v32", nlayers=5)
diff --git a/scripts/models/deepseek-v32-5layer.sh b/scripts/models/deepseek-v32-5layer.sh
deleted file mode 100644
index 2466640afd5..00000000000
--- a/scripts/models/deepseek-v32-5layer.sh
+++ /dev/null
@@ -1 +0,0 @@
-MODEL_ARGS_NUM_LAYERS=5 source "$(dirname -- "${BASH_SOURCE[0]}")/deepseek-v32.sh"
diff --git a/scripts/models/deepseek-v32.py b/scripts/models/deepseek-v32.py
new file mode 100644
index 00000000000..d2e02f006dc
--- /dev/null
+++ b/scripts/models/deepseek-v32.py
@@ -0,0 +1,58 @@
+import os
+
+from model_args_utils import moe_layer_freq
+
+
+def model_args(nlayers: int | None = None, first_k_dense_replace: int | None = None) -> str:
+    nlayers = nlayers if nlayers is not None else int(os.environ.get("MODEL_ARGS_NUM_LAYERS") or 61)
+    first_k_dense_replace = (
+        first_k_dense_replace
+        if first_k_dense_replace is not None
+        else int(os.environ.get("MODEL_ARGS_FIRST_K_DENSE_REPLACE") or 3)
+    )
+    return (
+        "--spec miles_plugins.models.glm5.glm5 get_glm5_spec "
+        "--disable-bias-linear "
+        f"--num-layers {nlayers} "
+        "--hidden-size 7168 "
+        "--ffn-hidden-size 18432 "
+        "--num-attention-heads 128 "
+        "--kv-channels 128 "
+        "--normalization RMSNorm "
+        "--position-embedding-type rope "
+        "--norm-epsilon 1e-6 "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--vocab-size 129280 "
+        "--multi-latent-attention "
+        "--q-lora-rank 1536 "
+        "--kv-lora-rank 512 "
+        "--qk-head-dim 128 "
+        "--qk-pos-emb-head-dim 64 "
+        "--v-head-dim 128 "
+        "--qk-layernorm "
+        "--rotary-scaling-factor 40 "
+        "--rotary-base 10000 "
+        "--mscale 1.0 "
+        "--mscale-all-dim 1.0 "
+        "--attention-softmax-in-fp32 "
+        "--no-rope-fusion "
+        "--num-experts 256 "
+        f"--moe-layer-freq {moe_layer_freq(nlayers=nlayers, first_k_dense_replace=first_k_dense_replace)} "
+        "--moe-ffn-hidden-size 2048 "
+        "--moe-router-topk 8 "
+        "--moe-shared-expert-intermediate-size 2048 "
+        "--moe-router-pre-softmax "
+        "--moe-router-score-function sigmoid "
+        "--moe-router-enable-expert-bias "
+        "--moe-router-load-balancing-type seq_aux_loss "
+        "--moe-token-dispatcher-type alltoall "
+        "--moe-aux-loss-coeff 0 "
+        "--moe-router-bias-update-rate 0 "
+        "--moe-router-group-topk 4 "
+        "--moe-router-num-groups 8 "
+        "--moe-grouped-gemm "
+        "--moe-router-topk-scaling-factor 2.5 "
+        "--moe-router-dtype fp32 "
+        "--moe-permute-fusion "
+    )
diff --git a/scripts/models/deepseek-v32.sh b/scripts/models/deepseek-v32.sh
deleted file mode 100644
index 649c3f6baca..00000000000
--- a/scripts/models/deepseek-v32.sh
+++ /dev/null
@@ -1,62 +0,0 @@
-NLAYERS="${MODEL_ARGS_NUM_LAYERS:-61}"
-FIRST_K_DENSE_REPLACE="${MODEL_ARGS_FIRST_K_DENSE_REPLACE:-3}"
-
-arr=()
-for ((i=0; i<NLAYERS; i++)); do
-  if (( i < FIRST_K_DENSE_REPLACE )); then
-    arr+=(0)
-  else
-    arr+=(1)
-  fi
-done
-
-printf -v MOE_LAYER_FREQ "[%s]" "$(IFS=', '; echo "${arr[*]}")"
-
-MODEL_ARGS=(
-    --spec "miles_plugins.models.glm5.glm5" "get_glm5_spec"
-    --disable-bias-linear
-    --num-layers $NLAYERS
-    --hidden-size 7168
-    --ffn-hidden-size 18432
-    --num-attention-heads 128
-    --kv-channels 128
-    --normalization RMSNorm
-    --position-embedding-type rope
-    --norm-epsilon 1e-6
-    --swiglu
-    --untie-embeddings-and-output-weights
-    --vocab-size 129280
-
-    --multi-latent-attention
-    --q-lora-rank 1536
-    --kv-lora-rank 512
-    --qk-head-dim 128
-    --qk-pos-emb-head-dim 64
-    --v-head-dim 128
-    --qk-layernorm
-    --rotary-scaling-factor 40
-    --rotary-base 10000
-    --mscale 1.0
-    --mscale-all-dim 1.0
-    --attention-softmax-in-fp32
-    --no-rope-fusion
-
-    --num-experts 256
-    --moe-layer-freq $MOE_LAYER_FREQ
-    --moe-ffn-hidden-size 2048
-    --moe-router-topk 8
-    --moe-shared-expert-intermediate-size 2048
-    --moe-router-pre-softmax
-    --moe-router-score-function sigmoid
-    --moe-router-enable-expert-bias
-    --moe-router-load-balancing-type seq_aux_loss
-    --moe-token-dispatcher-type alltoall
-    --moe-aux-loss-coeff 0
-    --moe-router-bias-update-rate 0
-    --moe-router-group-topk 4
-    --moe-router-num-groups 8
-    --moe-grouped-gemm
-    --moe-router-topk-scaling-factor 2.5
-    --moe-router-dtype fp32
-    --moe-permute-fusion
-)
diff --git a/scripts/models/deepseek-v4-flash-4layer.py b/scripts/models/deepseek-v4-flash-4layer.py
new file mode 100644
index 00000000000..2aff0e67ac8
--- /dev/null
+++ b/scripts/models/deepseek-v4-flash-4layer.py
@@ -0,0 +1,5 @@
+from model_args_utils import load_sibling_model_args
+
+
+def model_args() -> str:
+    return load_sibling_model_args(__file__, "deepseek-v4-flash", nlayers=4, compress_ratios="0 0 4 128")
diff --git a/scripts/models/deepseek-v4-flash-4layer.sh b/scripts/models/deepseek-v4-flash-4layer.sh
deleted file mode 100644
index 2a852460c1d..00000000000
--- a/scripts/models/deepseek-v4-flash-4layer.sh
+++ /dev/null
@@ -1,3 +0,0 @@
-MODEL_ARGS_NUM_LAYERS=4
-COMPRESS_RATIOS=(0 0 4 128)
-source "$(dirname -- "${BASH_SOURCE[0]}")/deepseek-v4-flash.sh"
diff --git a/scripts/models/deepseek-v4-flash.py b/scripts/models/deepseek-v4-flash.py
new file mode 100644
index 00000000000..f9b87961e3b
--- /dev/null
+++ b/scripts/models/deepseek-v4-flash.py
@@ -0,0 +1,77 @@
+import os
+
+from model_args_utils import moe_layer_freq
+
+
+COMPRESS_RATIOS = "0 0 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 0"
+SWIGLU_LIMIT_ARGS = "--activation-func-clamp-value 10 --no-bias-swiglu-fusion --no-activation-func-clamp-shared-expert"
+
+
+def model_args(
+    nlayers: int | None = None, rotary_scaling_factor: str | None = None, compress_ratios: str = COMPRESS_RATIOS
+) -> str:
+    nlayers = nlayers if nlayers is not None else int(os.environ.get("MODEL_ARGS_NUM_LAYERS") or 43)
+    rotary_scaling_factor = (
+        rotary_scaling_factor if rotary_scaling_factor is not None else os.environ.get("ROTARY_SCALING_FACTOR") or "16"
+    )
+    return (
+        "--disable-bias-linear "
+        f"--num-layers {nlayers} "
+        "--hidden-size 4096 "
+        "--ffn-hidden-size 2048 "
+        "--num-attention-heads 64 "
+        "--normalization RMSNorm "
+        "--position-embedding-type rope "
+        "--norm-epsilon 1e-6 "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--vocab-size 129280 "
+        "--hidden-dropout 0.0 "
+        "--attention-dropout 0.0 "
+        # MLA params (reused by V4)
+        "--multi-latent-attention "
+        "--q-lora-rank 1024 "
+        "--kv-lora-rank 512 "
+        "--qk-head-dim 512 "
+        "--qk-pos-emb-head-dim 64 "
+        "--v-head-dim 512 "
+        "--qk-layernorm "
+        f"--rotary-scaling-factor {rotary_scaling_factor} "
+        "--rotary-base 10000 "
+        "--original-max-position-embeddings 65536 "
+        "--beta-fast 32 "
+        "--beta-slow 1 "
+        "--attention-softmax-in-fp32 "
+        "--no-rope-fusion "
+        # MoE
+        "--num-experts 256 "
+        f"--moe-layer-freq {moe_layer_freq(nlayers=nlayers, first_k_dense_replace=0)} "
+        "--moe-ffn-hidden-size 2048 "
+        "--moe-router-topk 6 "
+        "--moe-shared-expert-intermediate-size 2048 "
+        "--moe-router-pre-softmax "
+        "--moe-router-score-function sqrtsoftplus "
+        "--moe-router-enable-expert-bias "
+        "--moe-router-load-balancing-type seq_aux_loss "
+        "--moe-token-dispatcher-type alltoall "
+        "--moe-aux-loss-coeff 0 "
+        "--moe-grouped-gemm "
+        "--moe-router-topk-scaling-factor 1.5 "
+        # DSV4 specific
+        "--experimental-attention-variant dsv4 "
+        "--dsv4-hc-mult 4 "
+        "--dsv4-hc-sinkhorn-iters 20 "
+        f"--dsv4-compress-ratios {compress_ratios} "
+        "--dsv4-compress-rope-theta 160000 "
+        "--dsv4-o-groups 8 "
+        "--dsv4-o-lora-rank 1024 "
+        "--dsv4-n-hash-layers 3 "
+        "--dsv4-window-size 128 "
+        # DSA Indexer
+        "--dsa-indexer-n-heads 64 "
+        "--dsa-indexer-head-dim 128 "
+        "--dsa-indexer-topk 512 "
+        # V4 model spec (plugin)
+        "--spec miles_plugins.models.deepseek_v4.deepseek_v4 get_dsv4_spec "
+        f"{SWIGLU_LIMIT_ARGS} "
+    )
diff --git a/scripts/models/deepseek-v4-flash.sh b/scripts/models/deepseek-v4-flash.sh
deleted file mode 100644
index 3997290de44..00000000000
--- a/scripts/models/deepseek-v4-flash.sh
+++ /dev/null
@@ -1,85 +0,0 @@
-NLAYERS="${MODEL_ARGS_NUM_LAYERS:-43}"
-
-# V4: all layers are MoE
-arr=()
-for ((i=0; i<NLAYERS; i++)); do
-  arr+=(1)
-done
-
-printf -v MOE_LAYER_FREQ "[%s]" "$(IFS=', '; echo "${arr[*]}")"
-
-if [ ${#COMPRESS_RATIOS[@]} -eq 0 ]; then
-  COMPRESS_RATIOS=(0 0 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 0)
-fi
-ROTARY_SCALING_FACTOR="${ROTARY_SCALING_FACTOR:-16}"
-
-SWIGLU_LIMIT_ARGS=(--activation-func-clamp-value 10 --no-bias-swiglu-fusion --no-activation-func-clamp-shared-expert)
-
-# DeepSeek V4 Flash config
-MODEL_ARGS=(
-    --disable-bias-linear
-    --num-layers $NLAYERS
-    --hidden-size 4096
-    --ffn-hidden-size 2048
-    --num-attention-heads 64
-    --normalization RMSNorm
-    --position-embedding-type rope
-    --norm-epsilon 1e-6
-    --swiglu
-    --untie-embeddings-and-output-weights
-    --vocab-size 129280
-    --hidden-dropout 0.0
-    --attention-dropout 0.0
-
-    # MLA params (reused by V4)
-    --multi-latent-attention
-    --q-lora-rank 1024
-    --kv-lora-rank 512
-    --qk-head-dim 512
-    --qk-pos-emb-head-dim 64
-    --v-head-dim 512
-    --qk-layernorm
-    --rotary-scaling-factor $ROTARY_SCALING_FACTOR
-    --rotary-base 10000
-    --original-max-position-embeddings 65536
-    --beta-fast 32
-    --beta-slow 1
-    --attention-softmax-in-fp32
-    --no-rope-fusion
-
-    # MoE
-    --num-experts 256
-    --moe-layer-freq $MOE_LAYER_FREQ
-    --moe-ffn-hidden-size 2048
-    --moe-router-topk 6
-    --moe-shared-expert-intermediate-size 2048
-    --moe-router-pre-softmax
-    --moe-router-score-function sqrtsoftplus
-    --moe-router-enable-expert-bias
-    --moe-router-load-balancing-type seq_aux_loss
-    --moe-token-dispatcher-type alltoall
-    --moe-aux-loss-coeff 0
-    --moe-grouped-gemm
-    --moe-router-topk-scaling-factor 1.5
-
-    # DSV4 specific
-    --experimental-attention-variant dsv4
-    --dsv4-hc-mult 4
-    --dsv4-hc-sinkhorn-iters 20
-    --dsv4-compress-ratios "${COMPRESS_RATIOS[@]}"
-    --dsv4-compress-rope-theta 160000
-    --dsv4-o-groups 8
-    --dsv4-o-lora-rank 1024
-    --dsv4-n-hash-layers 3
-    --dsv4-window-size 128
-
-    # DSA Indexer
-    --dsa-indexer-n-heads 64
-    --dsa-indexer-head-dim 128
-    --dsa-indexer-topk 512
-
-    # V4 model spec (plugin)
-    --spec miles_plugins.models.deepseek_v4.deepseek_v4 get_dsv4_spec
-
-    "${SWIGLU_LIMIT_ARGS[@]}"
-)
diff --git a/scripts/models/deepseek-v4-pro.py b/scripts/models/deepseek-v4-pro.py
new file mode 100644
index 00000000000..e2c8d44b2a7
--- /dev/null
+++ b/scripts/models/deepseek-v4-pro.py
@@ -0,0 +1,77 @@
+import os
+
+from model_args_utils import moe_layer_freq
+
+
+COMPRESS_RATIOS = "128 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 0"
+SWIGLU_LIMIT_ARGS = "--activation-func-clamp-value 10 --no-bias-swiglu-fusion --no-activation-func-clamp-shared-expert"
+
+
+def model_args(
+    nlayers: int | None = None, rotary_scaling_factor: str | None = None, compress_ratios: str = COMPRESS_RATIOS
+) -> str:
+    nlayers = nlayers if nlayers is not None else int(os.environ.get("MODEL_ARGS_NUM_LAYERS") or 61)
+    rotary_scaling_factor = (
+        rotary_scaling_factor if rotary_scaling_factor is not None else os.environ.get("ROTARY_SCALING_FACTOR") or "16"
+    )
+    return (
+        "--disable-bias-linear "
+        f"--num-layers {nlayers} "
+        "--hidden-size 7168 "
+        "--ffn-hidden-size 3072 "
+        "--num-attention-heads 128 "
+        "--normalization RMSNorm "
+        "--position-embedding-type rope "
+        "--norm-epsilon 1e-6 "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--vocab-size 129280 "
+        "--hidden-dropout 0.0 "
+        "--attention-dropout 0.0 "
+        # MLA params (reused by V4)
+        "--multi-latent-attention "
+        "--q-lora-rank 1536 "
+        "--kv-lora-rank 512 "
+        "--qk-head-dim 512 "
+        "--qk-pos-emb-head-dim 64 "
+        "--v-head-dim 512 "
+        "--qk-layernorm "
+        f"--rotary-scaling-factor {rotary_scaling_factor} "
+        "--rotary-base 10000 "
+        "--original-max-position-embeddings 65536 "
+        "--beta-fast 32 "
+        "--beta-slow 1 "
+        "--attention-softmax-in-fp32 "
+        "--no-rope-fusion "
+        # MoE
+        "--num-experts 384 "
+        f"--moe-layer-freq {moe_layer_freq(nlayers=nlayers, first_k_dense_replace=0)} "
+        "--moe-ffn-hidden-size 3072 "
+        "--moe-router-topk 6 "
+        "--moe-shared-expert-intermediate-size 3072 "
+        "--moe-router-pre-softmax "
+        "--moe-router-score-function sqrtsoftplus "
+        "--moe-router-enable-expert-bias "
+        "--moe-router-load-balancing-type seq_aux_loss "
+        "--moe-token-dispatcher-type alltoall "
+        "--moe-aux-loss-coeff 0 "
+        "--moe-grouped-gemm "
+        "--moe-router-topk-scaling-factor 2.5 "
+        # DSV4 specific
+        "--experimental-attention-variant dsv4 "
+        "--dsv4-hc-mult 4 "
+        "--dsv4-hc-sinkhorn-iters 20 "
+        f"--dsv4-compress-ratios {compress_ratios} "
+        "--dsv4-compress-rope-theta 160000 "
+        "--dsv4-o-groups 16 "
+        "--dsv4-o-lora-rank 1024 "
+        "--dsv4-n-hash-layers 3 "
+        "--dsv4-window-size 128 "
+        # DSA Indexer
+        "--dsa-indexer-n-heads 64 "
+        "--dsa-indexer-head-dim 128 "
+        "--dsa-indexer-topk 1024 "
+        # V4 model spec (plugin)
+        "--spec miles_plugins.models.deepseek_v4.deepseek_v4 get_dsv4_spec "
+        f"{SWIGLU_LIMIT_ARGS} "
+    )
diff --git a/scripts/models/deepseek-v4-pro.sh b/scripts/models/deepseek-v4-pro.sh
deleted file mode 100644
index 067d887d967..00000000000
--- a/scripts/models/deepseek-v4-pro.sh
+++ /dev/null
@@ -1,85 +0,0 @@
-NLAYERS="${MODEL_ARGS_NUM_LAYERS:-61}"
-
-# V4: all layers are MoE
-arr=()
-for ((i=0; i<NLAYERS; i++)); do
-  arr+=(1)
-done
-
-printf -v MOE_LAYER_FREQ "[%s]" "$(IFS=', '; echo "${arr[*]}")"
-
-if [ ${#COMPRESS_RATIOS[@]} -eq 0 ]; then
-  COMPRESS_RATIOS=(128 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 128 4 0)
-fi
-ROTARY_SCALING_FACTOR="${ROTARY_SCALING_FACTOR:-16}"
-
-SWIGLU_LIMIT_ARGS=(--activation-func-clamp-value 10 --no-bias-swiglu-fusion --no-activation-func-clamp-shared-expert)
-
-# DeepSeek V4 Pro config
-MODEL_ARGS=(
-    --disable-bias-linear
-    --num-layers $NLAYERS
-    --hidden-size 7168
-    --ffn-hidden-size 3072
-    --num-attention-heads 128
-    --normalization RMSNorm
-    --position-embedding-type rope
-    --norm-epsilon 1e-6
-    --swiglu
-    --untie-embeddings-and-output-weights
-    --vocab-size 129280
-    --hidden-dropout 0.0
-    --attention-dropout 0.0
-
-    # MLA params (reused by V4)
-    --multi-latent-attention
-    --q-lora-rank 1536
-    --kv-lora-rank 512
-    --qk-head-dim 512
-    --qk-pos-emb-head-dim 64
-    --v-head-dim 512
-    --qk-layernorm
-    --rotary-scaling-factor $ROTARY_SCALING_FACTOR
-    --rotary-base 10000
-    --original-max-position-embeddings 65536
-    --beta-fast 32
-    --beta-slow 1
-    --attention-softmax-in-fp32
-    --no-rope-fusion
-
-    # MoE
-    --num-experts 384
-    --moe-layer-freq $MOE_LAYER_FREQ
-    --moe-ffn-hidden-size 3072
-    --moe-router-topk 6
-    --moe-shared-expert-intermediate-size 3072
-    --moe-router-pre-softmax
-    --moe-router-score-function sqrtsoftplus
-    --moe-router-enable-expert-bias
-    --moe-router-load-balancing-type seq_aux_loss
-    --moe-token-dispatcher-type alltoall
-    --moe-aux-loss-coeff 0
-    --moe-grouped-gemm
-    --moe-router-topk-scaling-factor 2.5
-
-    # DSV4 specific
-    --experimental-attention-variant dsv4
-    --dsv4-hc-mult 4
-    --dsv4-hc-sinkhorn-iters 20
-    --dsv4-compress-ratios "${COMPRESS_RATIOS[@]}"
-    --dsv4-compress-rope-theta 160000
-    --dsv4-o-groups 16
-    --dsv4-o-lora-rank 1024
-    --dsv4-n-hash-layers 3
-    --dsv4-window-size 128
-
-    # DSA Indexer
-    --dsa-indexer-n-heads 64
-    --dsa-indexer-head-dim 128
-    --dsa-indexer-topk 1024
-
-    # V4 model spec (plugin)
-    --spec miles_plugins.models.deepseek_v4.deepseek_v4 get_dsv4_spec
-
-    "${SWIGLU_LIMIT_ARGS[@]}"
-)
diff --git a/scripts/models/gemma-4-26b-a4b-it.py b/scripts/models/gemma-4-26b-a4b-it.py
new file mode 100644
index 00000000000..a8189e8cd74
--- /dev/null
+++ b/scripts/models/gemma-4-26b-a4b-it.py
@@ -0,0 +1,30 @@
+def model_args() -> str:
+    return (
+        "--disable-bias-linear "
+        "--group-query-attention "
+        "--num-attention-heads 16 "
+        "--num-query-groups 8 "
+        "--kv-channels 256 "
+        "--num-layers 30 "
+        "--hidden-size 2816 "
+        "--ffn-hidden-size 2112 "
+        "--normalization RMSNorm "
+        "--norm-epsilon 1e-06 "
+        "--position-embedding-type rope "
+        "--rotary-base 1000000 "
+        "--vocab-size 262144 "
+        "--make-vocab-size-divisible-by 128 "
+        "--max-position-embeddings 262144 "
+        # tied embeddings: do not pass --untie-embeddings-and-output-weights
+        "--num-experts 128 "
+        "--moe-router-topk 8 "
+        "--moe-ffn-hidden-size 704 "
+        "--moe-router-score-function softmax "
+        "--moe-grouped-gemm "
+        "--moe-router-dtype fp32 "
+        "--moe-router-num-groups 1 "
+        "--moe-router-group-topk 1 "
+        "--moe-router-load-balancing-type seq_aux_loss "
+        "--moe-router-bias-update-rate 0 "
+        "--moe-aux-loss-coeff 0 "
+    )
diff --git a/scripts/models/gemma-4-26b-a4b-it.sh b/scripts/models/gemma-4-26b-a4b-it.sh
deleted file mode 100644
index 63fcee46545..00000000000
--- a/scripts/models/gemma-4-26b-a4b-it.sh
+++ /dev/null
@@ -1,32 +0,0 @@
-# Google Gemma-4 26B-A4B-it (BF16, MoE: 128 experts, top-k 8).
-
-MODEL_ARGS=(
-   --disable-bias-linear
-   --group-query-attention
-   --num-attention-heads 16
-   --num-query-groups 8
-   --kv-channels 256
-   --num-layers 30
-   --hidden-size 2816
-   --ffn-hidden-size 2112
-   --normalization RMSNorm
-   --norm-epsilon 1e-06
-   --position-embedding-type rope
-   --rotary-base 1000000
-   --vocab-size 262144
-   --make-vocab-size-divisible-by 128
-   --max-position-embeddings 262144
-   # tied embeddings: do not pass --untie-embeddings-and-output-weights
-
-   --num-experts 128
-   --moe-router-topk 8
-   --moe-ffn-hidden-size 704
-   --moe-router-score-function softmax
-   --moe-grouped-gemm
-   --moe-router-dtype fp32
-   --moe-router-num-groups 1
-   --moe-router-group-topk 1
-   --moe-router-load-balancing-type seq_aux_loss
-   --moe-router-bias-update-rate 0
-   --moe-aux-loss-coeff 0
-)
diff --git a/scripts/models/gemma-4-31b-it.py b/scripts/models/gemma-4-31b-it.py
new file mode 100644
index 00000000000..bc1ce66293d
--- /dev/null
+++ b/scripts/models/gemma-4-31b-it.py
@@ -0,0 +1,19 @@
+def model_args() -> str:
+    return (
+        "--disable-bias-linear "
+        "--group-query-attention "
+        "--num-attention-heads 32 "
+        "--num-query-groups 16 "
+        "--kv-channels 256 "
+        "--num-layers 60 "
+        "--hidden-size 5376 "
+        "--ffn-hidden-size 21504 "
+        "--normalization RMSNorm "
+        "--norm-epsilon 1e-06 "
+        "--position-embedding-type rope "
+        "--rotary-base 1000000 "
+        "--vocab-size 262144 "
+        "--make-vocab-size-divisible-by 128 "
+        "--max-position-embeddings 262144 "
+        # tied embeddings: do not pass --untie-embeddings-and-output-weights
+    )
diff --git a/scripts/models/gemma-4-31b-it.sh b/scripts/models/gemma-4-31b-it.sh
deleted file mode 100644
index 3a65ddb1f5f..00000000000
--- a/scripts/models/gemma-4-31b-it.sh
+++ /dev/null
@@ -1,20 +0,0 @@
-# Google Gemma-4 31B-it (BF16, DENSE — no experts).
-
-MODEL_ARGS=(
-   --disable-bias-linear
-   --group-query-attention
-   --num-attention-heads 32
-   --num-query-groups 16
-   --kv-channels 256
-   --num-layers 60
-   --hidden-size 5376
-   --ffn-hidden-size 21504
-   --normalization RMSNorm
-   --norm-epsilon 1e-06
-   --position-embedding-type rope
-   --rotary-base 1000000
-   --vocab-size 262144
-   --make-vocab-size-divisible-by 128
-   --max-position-embeddings 262144
-   # tied embeddings: do not pass --untie-embeddings-and-output-weights
-)
diff --git a/scripts/models/glm4-32B.py b/scripts/models/glm4-32B.py
new file mode 100644
index 00000000000..53613b2065d
--- /dev/null
+++ b/scripts/models/glm4-32B.py
@@ -0,0 +1,25 @@
+def model_args() -> str:
+    return (
+        "--spec miles_plugins.models.glm4 get_glm_spec "
+        "--swiglu "
+        "--num-layers 64 "
+        "--hidden-size 6144 "
+        "--ffn-hidden-size 23040 "
+        "--num-attention-heads 48 "
+        "--max-position-embeddings 32768 "
+        "--seq-length 32768 "
+        "--use-rotary-position-embeddings "
+        "--disable-bias-linear "
+        "--normalization RMSNorm "
+        "--norm-epsilon 1e-5 "
+        "--rotary-base 10000 "
+        "--group-query-attention "
+        "--num-query-groups 8 "
+        "--vocab-size 151552 "
+        "--post-self-attn-layernorm "
+        "--post-mlp-layernorm "
+        "--rotary-interleaved "
+        "--rotary-percent 0.5 "
+        "--no-rope-fusion "
+        "--untie-embeddings-and-output-weights "
+    )
diff --git a/scripts/models/glm4-32B.sh b/scripts/models/glm4-32B.sh
deleted file mode 100644
index 15cf273fc9a..00000000000
--- a/scripts/models/glm4-32B.sh
+++ /dev/null
@@ -1,24 +0,0 @@
-MODEL_ARGS=(
-   --spec "miles_plugins.models.glm4" "get_glm_spec"
-   --swiglu
-   --num-layers 64
-   --hidden-size 6144
-   --ffn-hidden-size 23040
-   --num-attention-heads 48
-   --max-position-embeddings 32768
-   --seq-length 32768
-   --use-rotary-position-embeddings
-   --disable-bias-linear
-   --normalization "RMSNorm"
-   --norm-epsilon 1e-5
-   --rotary-base 10000
-   --group-query-attention
-   --num-query-groups 8
-   --vocab-size 151552
-   --post-self-attn-layernorm
-   --post-mlp-layernorm
-   --rotary-interleaved
-   --rotary-percent 0.5
-   --no-rope-fusion
-   --untie-embeddings-and-output-weights
-)
\ No newline at end of file
diff --git a/scripts/models/glm4-9B.py b/scripts/models/glm4-9B.py
new file mode 100644
index 00000000000..0c320b91f6d
--- /dev/null
+++ b/scripts/models/glm4-9B.py
@@ -0,0 +1,24 @@
+def model_args() -> str:
+    return (
+        "--spec miles_plugins.models.glm4 get_glm_spec "
+        "--swiglu "
+        "--num-layers 40 "
+        "--hidden-size 4096 "
+        "--ffn-hidden-size 13696 "
+        "--num-attention-heads 32 "
+        "--group-query-attention "
+        "--num-query-groups 2 "
+        "--use-rotary-position-embeddings "
+        "--disable-bias-linear "
+        "--add-qkv-bias "
+        "--normalization RMSNorm "
+        "--norm-epsilon 1e-5 "
+        "--rotary-base 10000 "
+        "--vocab-size 151552 "
+        "--post-self-attn-layernorm "
+        "--post-mlp-layernorm "
+        "--rotary-interleaved "
+        "--rotary-percent 0.5 "
+        "--no-rope-fusion "
+        "--untie-embeddings-and-output-weights "
+    )
diff --git a/scripts/models/glm4-9B.sh b/scripts/models/glm4-9B.sh
deleted file mode 100644
index 652b579cf20..00000000000
--- a/scripts/models/glm4-9B.sh
+++ /dev/null
@@ -1,23 +0,0 @@
-MODEL_ARGS=(
-   --spec "miles_plugins.models.glm4" "get_glm_spec"
-   --swiglu
-   --num-layers 40
-   --hidden-size 4096
-   --ffn-hidden-size 13696
-   --num-attention-heads 32
-   --group-query-attention
-   --num-query-groups 2
-   --use-rotary-position-embeddings
-   --disable-bias-linear
-   --add-qkv-bias
-   --normalization "RMSNorm"
-   --norm-epsilon 1e-5
-   --rotary-base 10000
-   --vocab-size 151552
-   --post-self-attn-layernorm
-   --post-mlp-layernorm
-   --rotary-interleaved
-   --rotary-percent 0.5
-   --no-rope-fusion
-   --untie-embeddings-and-output-weights
-)
\ No newline at end of file
diff --git a/scripts/models/glm4.5-106B-A12B.py b/scripts/models/glm4.5-106B-A12B.py
new file mode 100644
index 00000000000..ba7617c98c8
--- /dev/null
+++ b/scripts/models/glm4.5-106B-A12B.py
@@ -0,0 +1,39 @@
+N_DENSE_LAYERS = 1
+N_MOE_LAYERS = 45
+
+
+def model_args() -> str:
+    return (
+        "--disable-bias-linear "
+        "--group-query-attention "
+        "--num-attention-heads 96 "
+        "--num-query-groups 8 "
+        "--kv-channels 128 "
+        f"--num-layers {N_DENSE_LAYERS + N_MOE_LAYERS} "
+        "--hidden-size 4096 "
+        "--ffn-hidden-size 10944 "
+        "--add-qkv-bias "
+        "--normalization RMSNorm "
+        "--position-embedding-type rope "
+        "--rotary-percent 0.5 "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--vocab-size 151552 "
+        "--rotary-base 1000000 "
+        # moe
+        "--moe-ffn-hidden-size 1408 "
+        "--moe-shared-expert-intermediate-size 1408 "
+        "--moe-router-pre-softmax "
+        "--moe-router-score-function sigmoid "
+        "--moe-router-enable-expert-bias "
+        "--moe-router-bias-update-rate 0 "
+        "--moe-router-load-balancing-type seq_aux_loss "
+        "--moe-token-dispatcher-type alltoall "
+        "--moe-router-topk 8 "
+        f"--moe-layer-freq [0]*{N_DENSE_LAYERS}+[1]*{N_MOE_LAYERS} "
+        "--num-experts 128 "
+        "--moe-grouped-gemm "
+        "--moe-router-dtype fp32 "
+        "--moe-permute-fusion "
+        "--moe-aux-loss-coeff 0 "
+    )
diff --git a/scripts/models/glm4.5-106B-A12B.sh b/scripts/models/glm4.5-106B-A12B.sh
deleted file mode 100644
index 8c244164336..00000000000
--- a/scripts/models/glm4.5-106B-A12B.sh
+++ /dev/null
@@ -1,40 +0,0 @@
-N_DENSE_LAYERS=1
-N_MOE_LAYERS=45
-
-# glm4.5-106B-A12B
-MODEL_ARGS=(
-    --disable-bias-linear
-    --group-query-attention
-    --num-attention-heads 96
-    --num-query-groups 8
-    --kv-channels 128
-    --num-layers $((N_DENSE_LAYERS + N_MOE_LAYERS))
-    --hidden-size 4096
-    --ffn-hidden-size 10944
-
-    --add-qkv-bias
-    --normalization RMSNorm
-    --position-embedding-type rope
-    --rotary-percent 0.5
-    --swiglu
-    --untie-embeddings-and-output-weights
-    --vocab-size 151552
-    --rotary-base 1000000
-
-    # moe
-    --moe-ffn-hidden-size 1408
-    --moe-shared-expert-intermediate-size 1408
-    --moe-router-pre-softmax
-    --moe-router-score-function sigmoid
-    --moe-router-enable-expert-bias
-    --moe-router-bias-update-rate 0
-    --moe-router-load-balancing-type seq_aux_loss
-    --moe-token-dispatcher-type alltoall
-    --moe-router-topk 8
-    --moe-layer-freq "[0]*$N_DENSE_LAYERS+[1]*$N_MOE_LAYERS"
-    --num-experts 128
-    --moe-grouped-gemm
-    --moe-router-dtype fp32
-    --moe-permute-fusion
-    --moe-aux-loss-coeff 0
-)
\ No newline at end of file
diff --git a/scripts/models/glm4.5-355B-A32B.py b/scripts/models/glm4.5-355B-A32B.py
new file mode 100644
index 00000000000..7c25722c55b
--- /dev/null
+++ b/scripts/models/glm4.5-355B-A32B.py
@@ -0,0 +1,41 @@
+N_DENSE_LAYERS = 3
+N_MOE_LAYERS = 89
+
+
+def model_args() -> str:
+    return (
+        "--disable-bias-linear "
+        "--qk-layernorm "
+        "--group-query-attention "
+        "--num-attention-heads 96 "
+        "--num-query-groups 8 "
+        "--kv-channels 128 "
+        f"--num-layers {N_DENSE_LAYERS + N_MOE_LAYERS} "
+        "--hidden-size 5120 "
+        "--ffn-hidden-size 12288 "
+        "--add-qkv-bias "
+        "--normalization RMSNorm "
+        "--position-embedding-type rope "
+        "--rotary-percent 0.5 "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--vocab-size 151552 "
+        "--rotary-base 1000000 "
+        # moe
+        "--moe-ffn-hidden-size 1536 "
+        "--moe-shared-expert-intermediate-size 1536 "
+        "--moe-router-pre-softmax "
+        "--moe-router-score-function sigmoid "
+        "--moe-router-enable-expert-bias "
+        "--moe-router-bias-update-rate 0 "
+        "--moe-router-load-balancing-type seq_aux_loss "
+        "--moe-token-dispatcher-type alltoall "
+        "--moe-router-topk 8 "
+        "--moe-router-topk-scaling-factor 2.5 "
+        f"--moe-layer-freq [0]*{N_DENSE_LAYERS}+[1]*{N_MOE_LAYERS} "
+        "--num-experts 160 "
+        "--moe-grouped-gemm "
+        "--moe-router-dtype fp32 "
+        "--moe-permute-fusion "
+        "--moe-aux-loss-coeff 0 "
+    )
diff --git a/scripts/models/glm4.5-355B-A32B.sh b/scripts/models/glm4.5-355B-A32B.sh
deleted file mode 100644
index 68b590ed4a7..00000000000
--- a/scripts/models/glm4.5-355B-A32B.sh
+++ /dev/null
@@ -1,43 +0,0 @@
-N_DENSE_LAYERS=3
-N_MOE_LAYERS=89
-
-# glm4.5-355B-A32B
-MODEL_ARGS=(
-    --disable-bias-linear
-    --qk-layernorm
-    --group-query-attention
-    --num-attention-heads 96
-    --num-query-groups 8
-    --kv-channels 128
-    --num-layers $((N_DENSE_LAYERS + N_MOE_LAYERS))
-    --hidden-size 5120
-    --ffn-hidden-size 12288
-    
-    --add-qkv-bias
-     --normalization RMSNorm
-    --position-embedding-type rope
-    --rotary-percent 0.5
-    --swiglu
-    --untie-embeddings-and-output-weights
-    --vocab-size 151552
-
-    --rotary-base 1000000
-
-    # moe
-    --moe-ffn-hidden-size 1536
-    --moe-shared-expert-intermediate-size 1536
-    --moe-router-pre-softmax
-    --moe-router-score-function sigmoid
-    --moe-router-enable-expert-bias
-    --moe-router-bias-update-rate 0
-    --moe-router-load-balancing-type seq_aux_loss
-    --moe-token-dispatcher-type alltoall
-    --moe-router-topk 8
-    --moe-router-topk-scaling-factor 2.5
-    --moe-layer-freq "[0]*$N_DENSE_LAYERS+[1]*$N_MOE_LAYERS"
-    --num-experts 160
-    --moe-grouped-gemm
-    --moe-router-dtype fp32
-    --moe-permute-fusion
-    --moe-aux-loss-coeff 0
-)
diff --git a/scripts/models/glm4.7-flash.py b/scripts/models/glm4.7-flash.py
new file mode 100644
index 00000000000..24363d671b5
--- /dev/null
+++ b/scripts/models/glm4.7-flash.py
@@ -0,0 +1,55 @@
+MOE_ROUTED_EXPERTS = 64
+MOE_ACTIVE_ROUTED_EXPERTS = 4
+MOE_SHARED_EXPERTS = 1
+NHIDDEN = 2048
+MOE_FFN_HIDDEN = 1536
+MOE_SHARED_EXPERT_INTERMEDIATE_SIZE = MOE_FFN_HIDDEN * MOE_SHARED_EXPERTS
+FFN_HIDDEN = 10240
+N_DENSE_LAYERS = 1
+N_MOE_LAYERS = 46
+NHEADS = 20
+
+
+def model_args() -> str:
+    return (
+        f"--moe-layer-freq [0]*{N_DENSE_LAYERS}+[1]*{N_MOE_LAYERS} "
+        f"--num-experts {MOE_ROUTED_EXPERTS} "
+        f"--moe-shared-expert-intermediate-size {MOE_SHARED_EXPERT_INTERMEDIATE_SIZE} "
+        f"--moe-router-topk {MOE_ACTIVE_ROUTED_EXPERTS} "
+        "--moe-grouped-gemm "
+        "--moe-permute-fusion "
+        f"--moe-ffn-hidden-size {MOE_FFN_HIDDEN} "
+        "--moe-router-score-function sigmoid "
+        "--moe-router-pre-softmax "
+        "--moe-router-enable-expert-bias "
+        "--moe-router-bias-update-rate 0 "
+        "--moe-router-load-balancing-type seq_aux_loss "
+        "--moe-router-topk-scaling-factor 1.8 "
+        "--moe-aux-loss-coeff 0 "
+        "--moe-router-dtype fp32 "
+        "--make-vocab-size-divisible-by 64 "
+        f"--num-layers {N_DENSE_LAYERS + N_MOE_LAYERS} "
+        f"--hidden-size {NHIDDEN} "
+        f"--ffn-hidden-size {FFN_HIDDEN} "
+        f"--num-attention-heads {NHEADS} "
+        "--disable-bias-linear "
+        "--add-qkv-bias "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--position-embedding-type rope "
+        "--no-position-embedding "
+        "--normalization RMSNorm "
+        "--norm-epsilon 1e-5 "
+        "--qk-layernorm "
+        "--multi-latent-attention "
+        "--q-lora-rank 768 "
+        "--kv-lora-rank 512 "
+        "--qk-head-dim 192 "
+        "--v-head-dim 256 "
+        "--kv-channels 192 "
+        "--qk-pos-emb-head-dim 64 "
+        "--vocab-size 154880 "
+        "--rotary-base 1000000 "
+        "--no-rope-fusion "
+        "--mtp-num-layers 1 "
+    )
diff --git a/scripts/models/glm4.7-flash.sh b/scripts/models/glm4.7-flash.sh
deleted file mode 100644
index 763aadc387a..00000000000
--- a/scripts/models/glm4.7-flash.sh
+++ /dev/null
@@ -1,54 +0,0 @@
-MOE_ROUTED_EXPERTS=64
-MOE_ACTIVE_ROUTED_EXPERTS=4
-MOE_SHARED_EXPERTS=1
-
-NHIDDEN=2048
-MOE_FFN_HIDDEN=1536
-MOE_SHARED_EXPERT_INTERMEDIATE_SIZE=$((MOE_FFN_HIDDEN * MOE_SHARED_EXPERTS))
-FFN_HIDDEN=10240
-N_DENSE_LAYERS=1
-N_MOE_LAYERS=46
-NHEADS=20
-
-MODEL_ARGS=(
-    --moe-layer-freq "[0]*${N_DENSE_LAYERS}+[1]*${N_MOE_LAYERS}"
-    --num-experts $MOE_ROUTED_EXPERTS
-    --moe-shared-expert-intermediate-size $MOE_SHARED_EXPERT_INTERMEDIATE_SIZE
-    --moe-router-topk $MOE_ACTIVE_ROUTED_EXPERTS
-    --moe-grouped-gemm
-    --moe-permute-fusion
-    --moe-ffn-hidden-size $MOE_FFN_HIDDEN
-    --moe-router-score-function sigmoid
-    --moe-router-pre-softmax
-    --moe-router-enable-expert-bias
-    --moe-router-bias-update-rate 0
-    --moe-router-load-balancing-type seq_aux_loss
-    --moe-router-topk-scaling-factor 1.8
-    --moe-aux-loss-coeff 0
-    --moe-router-dtype fp32
-    --make-vocab-size-divisible-by 64
-    --num-layers $((N_DENSE_LAYERS + N_MOE_LAYERS))
-    --hidden-size $NHIDDEN
-    --ffn-hidden-size $FFN_HIDDEN
-    --num-attention-heads $NHEADS
-    --disable-bias-linear
-    --add-qkv-bias
-    --swiglu
-    --untie-embeddings-and-output-weights
-    --position-embedding-type rope
-    --no-position-embedding
-    --normalization RMSNorm
-    --norm-epsilon 1e-5
-    --qk-layernorm
-    --multi-latent-attention
-    --q-lora-rank 768
-    --kv-lora-rank 512
-    --qk-head-dim 192
-    --v-head-dim 256
-    --kv-channels 192
-    --qk-pos-emb-head-dim 64
-    --vocab-size 154880
-    --rotary-base 1000000
-    --no-rope-fusion
-    --mtp-num-layers 1
-)
\ No newline at end of file
diff --git a/scripts/models/glm5-744B-A40B.py b/scripts/models/glm5-744B-A40B.py
new file mode 100644
index 00000000000..9869f411e8e
--- /dev/null
+++ b/scripts/models/glm5-744B-A40B.py
@@ -0,0 +1,52 @@
+MOE_ROUTED_EXPERTS = 256
+MOE_ACTIVE_ROUTED_EXPERTS = 8
+MOE_SHARED_EXPERTS = 1
+NHIDDEN = 6144
+MOE_FFN_HIDDEN = 2048
+MOE_SHARED_EXPERT_INTERMEDIATE_SIZE = MOE_FFN_HIDDEN * MOE_SHARED_EXPERTS
+FFN_HIDDEN = 12288
+N_DENSE_LAYERS = 3
+NHEADS = 64
+
+
+def model_args(n_moe_layers: int = 75) -> str:
+    return (
+        "--spec miles_plugins.models.glm5.glm5 get_glm5_spec "
+        f"--moe-layer-freq [0]*{N_DENSE_LAYERS}+[1]*{n_moe_layers} "
+        f"--num-experts {MOE_ROUTED_EXPERTS} "
+        f"--moe-shared-expert-intermediate-size {MOE_SHARED_EXPERT_INTERMEDIATE_SIZE} "
+        f"--moe-router-topk {MOE_ACTIVE_ROUTED_EXPERTS} "
+        "--moe-grouped-gemm "
+        "--moe-permute-fusion "
+        f"--moe-ffn-hidden-size {MOE_FFN_HIDDEN} "
+        "--moe-router-score-function sigmoid "
+        "--moe-router-pre-softmax "
+        "--moe-router-enable-expert-bias "
+        "--moe-router-bias-update-rate 0 "
+        "--moe-router-load-balancing-type seq_aux_loss "
+        "--moe-router-topk-scaling-factor 2.5 "
+        "--moe-aux-loss-coeff 0 "
+        "--moe-router-dtype fp32 "
+        "--make-vocab-size-divisible-by 16 "
+        f"--num-layers {N_DENSE_LAYERS + n_moe_layers} "
+        f"--hidden-size {NHIDDEN} "
+        f"--ffn-hidden-size {FFN_HIDDEN} "
+        f"--num-attention-heads {NHEADS} "
+        "--disable-bias-linear "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--position-embedding-type rope "
+        "--no-position-embedding "
+        "--normalization RMSNorm "
+        "--qk-layernorm "
+        "--multi-latent-attention "
+        "--q-lora-rank 2048 "
+        "--kv-lora-rank 512 "
+        "--qk-head-dim 192 "
+        "--v-head-dim 256 "
+        "--kv-channels 192 "
+        "--qk-pos-emb-head-dim 64 "
+        "--vocab-size 154880 "
+        "--rotary-base 1000000 "
+        "--enable-experimental "
+    )
diff --git a/scripts/models/glm5-744B-A40B.sh b/scripts/models/glm5-744B-A40B.sh
deleted file mode 100644
index b9241e1a09c..00000000000
--- a/scripts/models/glm5-744B-A40B.sh
+++ /dev/null
@@ -1,52 +0,0 @@
-MOE_ROUTED_EXPERTS=256
-MOE_ACTIVE_ROUTED_EXPERTS=8
-MOE_SHARED_EXPERTS=1
-
-NHIDDEN=6144
-MOE_FFN_HIDDEN=2048
-MOE_SHARED_EXPERT_INTERMEDIATE_SIZE=$(($MOE_FFN_HIDDEN * $MOE_SHARED_EXPERTS))
-FFN_HIDDEN=12288
-N_DENSE_LAYERS=3
-N_MOE_LAYERS=75
-NHEADS=64
-
-MODEL_ARGS=(
-   --spec "miles_plugins.models.glm5.glm5" "get_glm5_spec"
-    --moe-layer-freq "[0]*${N_DENSE_LAYERS}+[1]*${N_MOE_LAYERS}"
-    --num-experts $MOE_ROUTED_EXPERTS
-    --moe-shared-expert-intermediate-size $MOE_SHARED_EXPERT_INTERMEDIATE_SIZE
-    --moe-router-topk $MOE_ACTIVE_ROUTED_EXPERTS
-    --moe-grouped-gemm
-    --moe-permute-fusion
-    --moe-ffn-hidden-size $MOE_FFN_HIDDEN
-    --moe-router-score-function sigmoid
-    --moe-router-pre-softmax
-    --moe-router-enable-expert-bias
-    --moe-router-bias-update-rate 0
-    --moe-router-load-balancing-type seq_aux_loss
-    --moe-router-topk-scaling-factor 2.5
-    --moe-aux-loss-coeff 0
-    --moe-router-dtype fp32
-    --make-vocab-size-divisible-by 16
-    --num-layers $((N_DENSE_LAYERS + N_MOE_LAYERS))
-    --hidden-size $NHIDDEN
-    --ffn-hidden-size $FFN_HIDDEN
-    --num-attention-heads $NHEADS
-    --disable-bias-linear
-    --swiglu
-    --untie-embeddings-and-output-weights
-    --position-embedding-type rope
-    --no-position-embedding
-    --normalization RMSNorm
-    --qk-layernorm
-    --multi-latent-attention
-    --q-lora-rank 2048
-    --kv-lora-rank 512
-    --qk-head-dim 192
-    --v-head-dim 256
-    --kv-channels 192
-    --qk-pos-emb-head-dim 64
-    --vocab-size 154880
-    --rotary-base 1000000
-    --enable-experimental
-)
\ No newline at end of file
diff --git a/scripts/models/glm5-744B-A40B_20layer.py b/scripts/models/glm5-744B-A40B_20layer.py
new file mode 100644
index 00000000000..6ec017c9810
--- /dev/null
+++ b/scripts/models/glm5-744B-A40B_20layer.py
@@ -0,0 +1,6 @@
+from model_args_utils import load_sibling_model_args
+
+
+def model_args() -> str:
+    # Override for 20-layer pruned model (first 20 layers: 3 dense + 17 MoE)
+    return load_sibling_model_args(__file__, "glm5-744B-A40B", n_moe_layers=17)
diff --git a/scripts/models/glm5-744B-A40B_20layer.sh b/scripts/models/glm5-744B-A40B_20layer.sh
deleted file mode 100644
index 1eb85223d71..00000000000
--- a/scripts/models/glm5-744B-A40B_20layer.sh
+++ /dev/null
@@ -1,12 +0,0 @@
-SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]:-$0}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/glm5-744B-A40B.sh"
-
-# Override for 20-layer pruned model (first 20 layers: 3 dense + 17 MoE)
-N_MOE_LAYERS=17
-
-for ((i=0; i<${#MODEL_ARGS[@]}; i++)); do
-    case "${MODEL_ARGS[$i]}" in
-        --num-layers) MODEL_ARGS[$((i+1))]=$((N_DENSE_LAYERS + N_MOE_LAYERS)) ;;
-        --moe-layer-freq) MODEL_ARGS[$((i+1))]="[0]*${N_DENSE_LAYERS}+[1]*${N_MOE_LAYERS}" ;;
-    esac
-done
diff --git a/scripts/models/glm5-744B-A40B_4layer.py b/scripts/models/glm5-744B-A40B_4layer.py
new file mode 100644
index 00000000000..0a5c3ce5625
--- /dev/null
+++ b/scripts/models/glm5-744B-A40B_4layer.py
@@ -0,0 +1,6 @@
+from model_args_utils import load_sibling_model_args
+
+
+def model_args() -> str:
+    # Override for 4-layer pruned model (first 4 layers: 3 dense + 1 MoE)
+    return load_sibling_model_args(__file__, "glm5-744B-A40B", n_moe_layers=1)
diff --git a/scripts/models/glm5-744B-A40B_4layer.sh b/scripts/models/glm5-744B-A40B_4layer.sh
deleted file mode 100644
index 50f10056327..00000000000
--- a/scripts/models/glm5-744B-A40B_4layer.sh
+++ /dev/null
@@ -1,12 +0,0 @@
-SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]:-$0}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/glm5-744B-A40B.sh"
-
-# Override for 4-layer pruned model (first 4 layers: 3 dense + 1 MoE)
-N_MOE_LAYERS=1
-
-for ((i=0; i<${#MODEL_ARGS[@]}; i++)); do
-    case "${MODEL_ARGS[$i]}" in
-        --num-layers) MODEL_ARGS[$((i+1))]=$((N_DENSE_LAYERS + N_MOE_LAYERS)) ;;
-        --moe-layer-freq) MODEL_ARGS[$((i+1))]="[0]*${N_DENSE_LAYERS}+[1]*${N_MOE_LAYERS}" ;;
-    esac
-done
diff --git a/scripts/models/glm5.1-744B-A40B_6layer.py b/scripts/models/glm5.1-744B-A40B_6layer.py
new file mode 100644
index 00000000000..7144d28349a
--- /dev/null
+++ b/scripts/models/glm5.1-744B-A40B_6layer.py
@@ -0,0 +1,7 @@
+from model_args_utils import load_sibling_model_args
+
+
+def model_args() -> str:
+    # Override for the 6-layer pruned GLM-5.1 toy (jybsuper/GLM-5.1-6layer):
+    # first 6 layers = 3 dense + 3 MoE.
+    return load_sibling_model_args(__file__, "glm5-744B-A40B", n_moe_layers=3)
diff --git a/scripts/models/glm5.1-744B-A40B_6layer.sh b/scripts/models/glm5.1-744B-A40B_6layer.sh
deleted file mode 100644
index a7044d3a439..00000000000
--- a/scripts/models/glm5.1-744B-A40B_6layer.sh
+++ /dev/null
@@ -1,14 +0,0 @@
-SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]:-$0}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/glm5-744B-A40B.sh"
-
-# Override for the 6-layer pruned GLM-5.1 toy (jybsuper/GLM-5.1-6layer):
-# first 6 layers = 3 dense + 3 MoE.
-N_MOE_LAYERS=3
-
-for ((i=0; i<${#MODEL_ARGS[@]}; i++)); do
-    case "${MODEL_ARGS[$i]}" in
-        --num-layers) MODEL_ARGS[$((i+1))]=$((N_DENSE_LAYERS + N_MOE_LAYERS)) ;;
-        --moe-layer-freq) MODEL_ARGS[$((i+1))]="[0]*${N_DENSE_LAYERS}+[1]*${N_MOE_LAYERS}" ;;
-    esac
-done
-
diff --git a/scripts/models/glm5.1-744B-A40B_6layer_lora.py b/scripts/models/glm5.1-744B-A40B_6layer_lora.py
new file mode 100644
index 00000000000..efc537a2ed8
--- /dev/null
+++ b/scripts/models/glm5.1-744B-A40B_6layer_lora.py
@@ -0,0 +1,7 @@
+from model_args_utils import load_sibling_model_args
+
+
+def model_args() -> str:
+    # Override for the 6-layer pruned GLM-5.1 toy (jybsuper/GLM-5.1-6layer):
+    # first 6 layers = 3 dense + 3 MoE.
+    return load_sibling_model_args(__file__, "glm5.1-744B-A40B_lora", n_moe_layers=3)
diff --git a/scripts/models/glm5.1-744B-A40B_6layer_lora.sh b/scripts/models/glm5.1-744B-A40B_6layer_lora.sh
deleted file mode 100644
index 2cac8891cac..00000000000
--- a/scripts/models/glm5.1-744B-A40B_6layer_lora.sh
+++ /dev/null
@@ -1,13 +0,0 @@
-SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]:-$0}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/glm5.1-744B-A40B_lora.sh"
-
-# Override for the 6-layer pruned GLM-5.1 toy (jybsuper/GLM-5.1-6layer):
-# first 6 layers = 3 dense + 3 MoE.
-N_MOE_LAYERS=3
-
-for ((i=0; i<${#MODEL_ARGS[@]}; i++)); do
-    case "${MODEL_ARGS[$i]}" in
-        --num-layers) MODEL_ARGS[$((i+1))]=$((N_DENSE_LAYERS + N_MOE_LAYERS)) ;;
-        --moe-layer-freq) MODEL_ARGS[$((i+1))]="[0]*${N_DENSE_LAYERS}+[1]*${N_MOE_LAYERS}" ;;
-    esac
-done
diff --git a/scripts/models/glm5.1-744B-A40B_lora.py b/scripts/models/glm5.1-744B-A40B_lora.py
new file mode 100644
index 00000000000..9869f411e8e
--- /dev/null
+++ b/scripts/models/glm5.1-744B-A40B_lora.py
@@ -0,0 +1,52 @@
+MOE_ROUTED_EXPERTS = 256
+MOE_ACTIVE_ROUTED_EXPERTS = 8
+MOE_SHARED_EXPERTS = 1
+NHIDDEN = 6144
+MOE_FFN_HIDDEN = 2048
+MOE_SHARED_EXPERT_INTERMEDIATE_SIZE = MOE_FFN_HIDDEN * MOE_SHARED_EXPERTS
+FFN_HIDDEN = 12288
+N_DENSE_LAYERS = 3
+NHEADS = 64
+
+
+def model_args(n_moe_layers: int = 75) -> str:
+    return (
+        "--spec miles_plugins.models.glm5.glm5 get_glm5_spec "
+        f"--moe-layer-freq [0]*{N_DENSE_LAYERS}+[1]*{n_moe_layers} "
+        f"--num-experts {MOE_ROUTED_EXPERTS} "
+        f"--moe-shared-expert-intermediate-size {MOE_SHARED_EXPERT_INTERMEDIATE_SIZE} "
+        f"--moe-router-topk {MOE_ACTIVE_ROUTED_EXPERTS} "
+        "--moe-grouped-gemm "
+        "--moe-permute-fusion "
+        f"--moe-ffn-hidden-size {MOE_FFN_HIDDEN} "
+        "--moe-router-score-function sigmoid "
+        "--moe-router-pre-softmax "
+        "--moe-router-enable-expert-bias "
+        "--moe-router-bias-update-rate 0 "
+        "--moe-router-load-balancing-type seq_aux_loss "
+        "--moe-router-topk-scaling-factor 2.5 "
+        "--moe-aux-loss-coeff 0 "
+        "--moe-router-dtype fp32 "
+        "--make-vocab-size-divisible-by 16 "
+        f"--num-layers {N_DENSE_LAYERS + n_moe_layers} "
+        f"--hidden-size {NHIDDEN} "
+        f"--ffn-hidden-size {FFN_HIDDEN} "
+        f"--num-attention-heads {NHEADS} "
+        "--disable-bias-linear "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--position-embedding-type rope "
+        "--no-position-embedding "
+        "--normalization RMSNorm "
+        "--qk-layernorm "
+        "--multi-latent-attention "
+        "--q-lora-rank 2048 "
+        "--kv-lora-rank 512 "
+        "--qk-head-dim 192 "
+        "--v-head-dim 256 "
+        "--kv-channels 192 "
+        "--qk-pos-emb-head-dim 64 "
+        "--vocab-size 154880 "
+        "--rotary-base 1000000 "
+        "--enable-experimental "
+    )
diff --git a/scripts/models/glm5.1-744B-A40B_lora.sh b/scripts/models/glm5.1-744B-A40B_lora.sh
deleted file mode 100644
index 2b77e7c71d7..00000000000
--- a/scripts/models/glm5.1-744B-A40B_lora.sh
+++ /dev/null
@@ -1,58 +0,0 @@
-MOE_ROUTED_EXPERTS=256
-MOE_ACTIVE_ROUTED_EXPERTS=8
-MOE_SHARED_EXPERTS=1
-
-NHIDDEN=6144
-MOE_FFN_HIDDEN=2048
-MOE_SHARED_EXPERT_INTERMEDIATE_SIZE=$(($MOE_FFN_HIDDEN * $MOE_SHARED_EXPERTS))
-FFN_HIDDEN=12288
-N_DENSE_LAYERS=3
-N_MOE_LAYERS=75
-NHEADS=64
-
-# GLM-5.1 744B-A40B (zai-org/GLM-5.1, glm_moe_dsa). Identical to glm5-744B-A40B.sh; the only
-# architecture difference vs the glm5.2-744B-A40B* registries is --rotary-base (5.2 uses 8e6).
-MODEL_ARGS=(
-   --spec "miles_plugins.models.glm5.glm5" "get_glm5_spec"
-    --moe-layer-freq "[0]*${N_DENSE_LAYERS}+[1]*${N_MOE_LAYERS}"
-    --num-experts $MOE_ROUTED_EXPERTS
-    --moe-shared-expert-intermediate-size $MOE_SHARED_EXPERT_INTERMEDIATE_SIZE
-    --moe-router-topk $MOE_ACTIVE_ROUTED_EXPERTS
-    --moe-grouped-gemm
-    --moe-permute-fusion
-    --moe-ffn-hidden-size $MOE_FFN_HIDDEN
-    --moe-router-score-function sigmoid
-    --moe-router-pre-softmax
-    --moe-router-enable-expert-bias
-    --moe-router-bias-update-rate 0
-    --moe-router-load-balancing-type seq_aux_loss
-    --moe-router-topk-scaling-factor 2.5
-    --moe-aux-loss-coeff 0
-    --moe-router-dtype fp32
-    --make-vocab-size-divisible-by 16
-    --num-layers $((N_DENSE_LAYERS + N_MOE_LAYERS))
-    --hidden-size $NHIDDEN
-    --ffn-hidden-size $FFN_HIDDEN
-    --num-attention-heads $NHEADS
-    --disable-bias-linear
-    --swiglu
-    --untie-embeddings-and-output-weights
-    --position-embedding-type rope
-    --no-position-embedding
-    --normalization RMSNorm
-    --qk-layernorm
-    --multi-latent-attention
-    --q-lora-rank 2048
-    --kv-lora-rank 512
-    --qk-head-dim 192
-    --v-head-dim 256
-    --kv-channels 192
-    --qk-pos-emb-head-dim 64
-    --vocab-size 154880
-    --rotary-base 1000000
-    --enable-experimental
-)
-
-# LoRA registry for scripts/run_glm5_1_744b_a40b_lora.py: MODEL_ARGS carries the architecture only
-# (identical to glm5-744B-A40B.sh); every LoRA / run-mode flag lives in the runner, which
-# always wins (argparse last-occurrence). --spec above is inert under bridge LoRA.
diff --git a/scripts/models/glm5.2-744B-A40B.py b/scripts/models/glm5.2-744B-A40B.py
new file mode 100644
index 00000000000..6f582e55594
--- /dev/null
+++ b/scripts/models/glm5.2-744B-A40B.py
@@ -0,0 +1,52 @@
+MOE_ROUTED_EXPERTS = 256
+MOE_ACTIVE_ROUTED_EXPERTS = 8
+MOE_SHARED_EXPERTS = 1
+NHIDDEN = 6144
+MOE_FFN_HIDDEN = 2048
+MOE_SHARED_EXPERT_INTERMEDIATE_SIZE = MOE_FFN_HIDDEN * MOE_SHARED_EXPERTS
+FFN_HIDDEN = 12288
+N_DENSE_LAYERS = 3
+NHEADS = 64
+
+
+def model_args(n_moe_layers: int = 75) -> str:
+    return (
+        "--spec miles_plugins.models.glm5.glm5 get_glm5_spec "
+        f"--moe-layer-freq [0]*{N_DENSE_LAYERS}+[1]*{n_moe_layers} "
+        f"--num-experts {MOE_ROUTED_EXPERTS} "
+        f"--moe-shared-expert-intermediate-size {MOE_SHARED_EXPERT_INTERMEDIATE_SIZE} "
+        f"--moe-router-topk {MOE_ACTIVE_ROUTED_EXPERTS} "
+        "--moe-grouped-gemm "
+        "--moe-permute-fusion "
+        f"--moe-ffn-hidden-size {MOE_FFN_HIDDEN} "
+        "--moe-router-score-function sigmoid "
+        "--moe-router-pre-softmax "
+        "--moe-router-enable-expert-bias "
+        "--moe-router-bias-update-rate 0 "
+        "--moe-router-load-balancing-type seq_aux_loss "
+        "--moe-router-topk-scaling-factor 2.5 "
+        "--moe-aux-loss-coeff 0 "
+        "--moe-router-dtype fp32 "
+        "--make-vocab-size-divisible-by 16 "
+        f"--num-layers {N_DENSE_LAYERS + n_moe_layers} "
+        f"--hidden-size {NHIDDEN} "
+        f"--ffn-hidden-size {FFN_HIDDEN} "
+        f"--num-attention-heads {NHEADS} "
+        "--disable-bias-linear "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--position-embedding-type rope "
+        "--no-position-embedding "
+        "--normalization RMSNorm "
+        "--qk-layernorm "
+        "--multi-latent-attention "
+        "--q-lora-rank 2048 "
+        "--kv-lora-rank 512 "
+        "--qk-head-dim 192 "
+        "--v-head-dim 256 "
+        "--kv-channels 192 "
+        "--qk-pos-emb-head-dim 64 "
+        "--vocab-size 154880 "
+        "--rotary-base 8000000 "
+        "--enable-experimental "
+    )
diff --git a/scripts/models/glm5.2-744B-A40B.sh b/scripts/models/glm5.2-744B-A40B.sh
deleted file mode 100644
index 2b831c6f185..00000000000
--- a/scripts/models/glm5.2-744B-A40B.sh
+++ /dev/null
@@ -1,59 +0,0 @@
-MOE_ROUTED_EXPERTS=256
-MOE_ACTIVE_ROUTED_EXPERTS=8
-MOE_SHARED_EXPERTS=1
-
-NHIDDEN=6144
-MOE_FFN_HIDDEN=2048
-MOE_SHARED_EXPERT_INTERMEDIATE_SIZE=$(($MOE_FFN_HIDDEN * $MOE_SHARED_EXPERTS))
-FFN_HIDDEN=12288
-N_DENSE_LAYERS=3
-N_MOE_LAYERS=75
-NHEADS=64
-
-# GLM-5.2 744B-A40B with DSA cross-layer index sharing. Only the computing layers
-# (1,2,3,7,11,...,75 in Megatron 1-indexing) carry indexer weights and compute the
-# sparse top-k; the remaining layers reuse the most recent computing layer's indices.
-# The schedule (index_topk_freq=4, index_skip_topk_offset=3) is read from the HF config
-# by the shared glm5 provider; cross-layer sharing activates when index_topk_freq > 1.
-# allgather-CP is enabled at train time in the run script (not here) so that checkpoint
-# conversion does not need to parse it. Differs from glm5-744B-A40B.sh only in rotary-base.
-MODEL_ARGS=(
-   --spec "miles_plugins.models.glm5.glm5" "get_glm5_spec"
-    --moe-layer-freq "[0]*${N_DENSE_LAYERS}+[1]*${N_MOE_LAYERS}"
-    --num-experts $MOE_ROUTED_EXPERTS
-    --moe-shared-expert-intermediate-size $MOE_SHARED_EXPERT_INTERMEDIATE_SIZE
-    --moe-router-topk $MOE_ACTIVE_ROUTED_EXPERTS
-    --moe-grouped-gemm
-    --moe-permute-fusion
-    --moe-ffn-hidden-size $MOE_FFN_HIDDEN
-    --moe-router-score-function sigmoid
-    --moe-router-pre-softmax
-    --moe-router-enable-expert-bias
-    --moe-router-bias-update-rate 0
-    --moe-router-load-balancing-type seq_aux_loss
-    --moe-router-topk-scaling-factor 2.5
-    --moe-aux-loss-coeff 0
-    --moe-router-dtype fp32
-    --make-vocab-size-divisible-by 16
-    --num-layers $((N_DENSE_LAYERS + N_MOE_LAYERS))
-    --hidden-size $NHIDDEN
-    --ffn-hidden-size $FFN_HIDDEN
-    --num-attention-heads $NHEADS
-    --disable-bias-linear
-    --swiglu
-    --untie-embeddings-and-output-weights
-    --position-embedding-type rope
-    --no-position-embedding
-    --normalization RMSNorm
-    --qk-layernorm
-    --multi-latent-attention
-    --q-lora-rank 2048
-    --kv-lora-rank 512
-    --qk-head-dim 192
-    --v-head-dim 256
-    --kv-channels 192
-    --qk-pos-emb-head-dim 64
-    --vocab-size 154880
-    --rotary-base 8000000
-    --enable-experimental
-)
diff --git a/scripts/models/glm5.2-744B-A40B_5layer.py b/scripts/models/glm5.2-744B-A40B_5layer.py
new file mode 100644
index 00000000000..35f03c3b9b6
--- /dev/null
+++ b/scripts/models/glm5.2-744B-A40B_5layer.py
@@ -0,0 +1,8 @@
+from model_args_utils import load_sibling_model_args
+
+
+def model_args() -> str:
+    # Override for 5-layer pruned model (first 5 layers: 3 dense + 2 MoE).
+    # Keeps at least one computing + one skip layer so the DSA cross-layer index
+    # sharing path is exercised (computing layers 0,1,2; skip layers 3,4).
+    return load_sibling_model_args(__file__, "glm5.2-744B-A40B", n_moe_layers=2)
diff --git a/scripts/models/glm5.2-744B-A40B_5layer.sh b/scripts/models/glm5.2-744B-A40B_5layer.sh
deleted file mode 100644
index f98c2c3ed9f..00000000000
--- a/scripts/models/glm5.2-744B-A40B_5layer.sh
+++ /dev/null
@@ -1,14 +0,0 @@
-SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]:-$0}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/glm5.2-744B-A40B.sh"
-
-# Override for 5-layer pruned model (first 5 layers: 3 dense + 2 MoE).
-# Keeps at least one computing + one skip layer so the DSA cross-layer index
-# sharing path is exercised (computing layers 0,1,2; skip layers 3,4).
-N_MOE_LAYERS=2
-
-for ((i=0; i<${#MODEL_ARGS[@]}; i++)); do
-    case "${MODEL_ARGS[$i]}" in
-        --num-layers) MODEL_ARGS[$((i+1))]=$((N_DENSE_LAYERS + N_MOE_LAYERS)) ;;
-        --moe-layer-freq) MODEL_ARGS[$((i+1))]="[0]*${N_DENSE_LAYERS}+[1]*${N_MOE_LAYERS}" ;;
-    esac
-done
diff --git a/scripts/models/glm5.2-744B-A40B_5layer_lora.py b/scripts/models/glm5.2-744B-A40B_5layer_lora.py
new file mode 100644
index 00000000000..0df3cf9e296
--- /dev/null
+++ b/scripts/models/glm5.2-744B-A40B_5layer_lora.py
@@ -0,0 +1,7 @@
+from model_args_utils import load_sibling_model_args
+
+
+def model_args() -> str:
+    # Override for the 5-layer pruned model (first 5 layers: 3 dense + 2 MoE). Keeps at least
+    # one computing + one skip layer so the DSA cross-layer index sharing path is exercised.
+    return load_sibling_model_args(__file__, "glm5.2-744B-A40B_lora", n_moe_layers=2)
diff --git a/scripts/models/glm5.2-744B-A40B_5layer_lora.sh b/scripts/models/glm5.2-744B-A40B_5layer_lora.sh
deleted file mode 100644
index 8da4e10f6ed..00000000000
--- a/scripts/models/glm5.2-744B-A40B_5layer_lora.sh
+++ /dev/null
@@ -1,13 +0,0 @@
-SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]:-$0}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/glm5.2-744B-A40B_lora.sh"
-
-# Override for the 5-layer pruned model (first 5 layers: 3 dense + 2 MoE). Keeps at least
-# one computing + one skip layer so the DSA cross-layer index sharing path is exercised.
-N_MOE_LAYERS=2
-
-for ((i=0; i<${#MODEL_ARGS[@]}; i++)); do
-    case "${MODEL_ARGS[$i]}" in
-        --num-layers) MODEL_ARGS[$((i+1))]=$((N_DENSE_LAYERS + N_MOE_LAYERS)) ;;
-        --moe-layer-freq) MODEL_ARGS[$((i+1))]="[0]*${N_DENSE_LAYERS}+[1]*${N_MOE_LAYERS}" ;;
-    esac
-done
diff --git a/scripts/models/glm5.2-744B-A40B_lora.py b/scripts/models/glm5.2-744B-A40B_lora.py
new file mode 100644
index 00000000000..6f582e55594
--- /dev/null
+++ b/scripts/models/glm5.2-744B-A40B_lora.py
@@ -0,0 +1,52 @@
+MOE_ROUTED_EXPERTS = 256
+MOE_ACTIVE_ROUTED_EXPERTS = 8
+MOE_SHARED_EXPERTS = 1
+NHIDDEN = 6144
+MOE_FFN_HIDDEN = 2048
+MOE_SHARED_EXPERT_INTERMEDIATE_SIZE = MOE_FFN_HIDDEN * MOE_SHARED_EXPERTS
+FFN_HIDDEN = 12288
+N_DENSE_LAYERS = 3
+NHEADS = 64
+
+
+def model_args(n_moe_layers: int = 75) -> str:
+    return (
+        "--spec miles_plugins.models.glm5.glm5 get_glm5_spec "
+        f"--moe-layer-freq [0]*{N_DENSE_LAYERS}+[1]*{n_moe_layers} "
+        f"--num-experts {MOE_ROUTED_EXPERTS} "
+        f"--moe-shared-expert-intermediate-size {MOE_SHARED_EXPERT_INTERMEDIATE_SIZE} "
+        f"--moe-router-topk {MOE_ACTIVE_ROUTED_EXPERTS} "
+        "--moe-grouped-gemm "
+        "--moe-permute-fusion "
+        f"--moe-ffn-hidden-size {MOE_FFN_HIDDEN} "
+        "--moe-router-score-function sigmoid "
+        "--moe-router-pre-softmax "
+        "--moe-router-enable-expert-bias "
+        "--moe-router-bias-update-rate 0 "
+        "--moe-router-load-balancing-type seq_aux_loss "
+        "--moe-router-topk-scaling-factor 2.5 "
+        "--moe-aux-loss-coeff 0 "
+        "--moe-router-dtype fp32 "
+        "--make-vocab-size-divisible-by 16 "
+        f"--num-layers {N_DENSE_LAYERS + n_moe_layers} "
+        f"--hidden-size {NHIDDEN} "
+        f"--ffn-hidden-size {FFN_HIDDEN} "
+        f"--num-attention-heads {NHEADS} "
+        "--disable-bias-linear "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--position-embedding-type rope "
+        "--no-position-embedding "
+        "--normalization RMSNorm "
+        "--qk-layernorm "
+        "--multi-latent-attention "
+        "--q-lora-rank 2048 "
+        "--kv-lora-rank 512 "
+        "--qk-head-dim 192 "
+        "--v-head-dim 256 "
+        "--kv-channels 192 "
+        "--qk-pos-emb-head-dim 64 "
+        "--vocab-size 154880 "
+        "--rotary-base 8000000 "
+        "--enable-experimental "
+    )
diff --git a/scripts/models/glm5.2-744B-A40B_lora.sh b/scripts/models/glm5.2-744B-A40B_lora.sh
deleted file mode 100644
index c9d2d4971f3..00000000000
--- a/scripts/models/glm5.2-744B-A40B_lora.sh
+++ /dev/null
@@ -1,58 +0,0 @@
-MOE_ROUTED_EXPERTS=256
-MOE_ACTIVE_ROUTED_EXPERTS=8
-MOE_SHARED_EXPERTS=1
-
-NHIDDEN=6144
-MOE_FFN_HIDDEN=2048
-MOE_SHARED_EXPERT_INTERMEDIATE_SIZE=$(($MOE_FFN_HIDDEN * $MOE_SHARED_EXPERTS))
-FFN_HIDDEN=12288
-N_DENSE_LAYERS=3
-N_MOE_LAYERS=75
-NHEADS=64
-
-# GLM-5.2 744B-A40B (DSA cross-layer index sharing; the schedule is read from the HF config
-# by the shared glm5 provider). Differs from glm5-744B-A40B.sh only in rotary-base.
-MODEL_ARGS=(
-   --spec "miles_plugins.models.glm5.glm5" "get_glm5_spec"
-    --moe-layer-freq "[0]*${N_DENSE_LAYERS}+[1]*${N_MOE_LAYERS}"
-    --num-experts $MOE_ROUTED_EXPERTS
-    --moe-shared-expert-intermediate-size $MOE_SHARED_EXPERT_INTERMEDIATE_SIZE
-    --moe-router-topk $MOE_ACTIVE_ROUTED_EXPERTS
-    --moe-grouped-gemm
-    --moe-permute-fusion
-    --moe-ffn-hidden-size $MOE_FFN_HIDDEN
-    --moe-router-score-function sigmoid
-    --moe-router-pre-softmax
-    --moe-router-enable-expert-bias
-    --moe-router-bias-update-rate 0
-    --moe-router-load-balancing-type seq_aux_loss
-    --moe-router-topk-scaling-factor 2.5
-    --moe-aux-loss-coeff 0
-    --moe-router-dtype fp32
-    --make-vocab-size-divisible-by 16
-    --num-layers $((N_DENSE_LAYERS + N_MOE_LAYERS))
-    --hidden-size $NHIDDEN
-    --ffn-hidden-size $FFN_HIDDEN
-    --num-attention-heads $NHEADS
-    --disable-bias-linear
-    --swiglu
-    --untie-embeddings-and-output-weights
-    --position-embedding-type rope
-    --no-position-embedding
-    --normalization RMSNorm
-    --qk-layernorm
-    --multi-latent-attention
-    --q-lora-rank 2048
-    --kv-lora-rank 512
-    --qk-head-dim 192
-    --v-head-dim 256
-    --kv-channels 192
-    --qk-pos-emb-head-dim 64
-    --vocab-size 154880
-    --rotary-base 8000000
-    --enable-experimental
-)
-
-# LoRA registry for scripts/run_glm5_2_744b_a40b_lora.py: MODEL_ARGS carries the architecture only
-# (identical to glm5.2-744B-A40B.sh); every LoRA / run-mode flag lives in the runner, which
-# always wins (argparse last-occurrence). --spec above is inert under bridge LoRA.
diff --git a/scripts/models/gpt-oss-20b.py b/scripts/models/gpt-oss-20b.py
new file mode 100644
index 00000000000..18009a85e96
--- /dev/null
+++ b/scripts/models/gpt-oss-20b.py
@@ -0,0 +1,40 @@
+def model_args() -> str:
+    return (
+        # Base architecture
+        "--num-layers 24 "
+        "--hidden-size 2880 "
+        "--ffn-hidden-size 2880 "
+        "--num-attention-heads 64 "
+        "--group-query-attention "
+        "--num-query-groups 8 "
+        "--kv-channels 64 "
+        # Positional embeddings
+        "--use-rotary-position-embeddings "
+        "--rotary-percent 1.0 "
+        "--rotary-base 150000 "
+        # Train with a 4k context, but keep max positions aligned with the HF checkpoint (YaRN scaling).
+        "--max-position-embeddings 131072 "
+        # Normalization
+        "--normalization RMSNorm "
+        "--norm-epsilon 1e-5 "
+        # Activation & embeddings
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--vocab-size 201088 "
+        # Note: attention_bias is true in HF config, so we may need bias
+        # --disable-bias-linear  # commented out since attention_bias=true
+        # Sliding window attention + learnable softmax offset (alternating SWA/full attention).
+        "--softmax-type learnable "
+        "--window-size 128,0 "
+        "--window-attn-skip-freq 2 "
+        # Fusions can be incompatible with this attention pattern on some stacks.
+        "--no-masked-softmax-fusion "
+        "--no-rope-fusion "
+        # MoE parameters
+        "--num-experts 32 "
+        "--moe-router-topk 4 "
+        "--moe-aux-loss-coeff 0.0 "
+        "--moe-token-dispatcher-type alltoall "
+        "--moe-router-dtype fp32 "
+        "--moe-grouped-gemm "
+    )
diff --git a/scripts/models/gpt-oss-20b.sh b/scripts/models/gpt-oss-20b.sh
deleted file mode 100644
index c1184f515ab..00000000000
--- a/scripts/models/gpt-oss-20b.sh
+++ /dev/null
@@ -1,48 +0,0 @@
-# gpt-oss-20b model architecture
-# Expected to match HF config for gpt-oss-20b-BF16 (MoE + sliding window attention).
-
-MODEL_ARGS=(
-   # Base architecture
-   --num-layers 24
-   --hidden-size 2880
-   --ffn-hidden-size 2880
-   --num-attention-heads 64
-   --group-query-attention
-   --num-query-groups 8
-   --kv-channels 64
-
-   # Positional embeddings
-   --use-rotary-position-embeddings
-   --rotary-percent 1.0
-   --rotary-base 150000
-   # Train with a 4k context, but keep max positions aligned with the HF checkpoint (YaRN scaling).
-   --max-position-embeddings 131072
-
-   # Normalization
-   --normalization "RMSNorm"
-   --norm-epsilon 1e-5
-
-   # Activation & embeddings
-   --swiglu
-   --untie-embeddings-and-output-weights
-   --vocab-size 201088
-
-   # Note: attention_bias is true in HF config, so we may need bias
-   # --disable-bias-linear  # commented out since attention_bias=true
-
-   # Sliding window attention + learnable softmax offset (alternating SWA/full attention).
-   --softmax-type learnable
-   --window-size 128,0
-   --window-attn-skip-freq 2
-   # Fusions can be incompatible with this attention pattern on some stacks.
-   --no-masked-softmax-fusion
-   --no-rope-fusion
-
-   # MoE parameters
-   --num-experts 32
-   --moe-router-topk 4
-   --moe-aux-loss-coeff 0.0
-   --moe-token-dispatcher-type alltoall
-   --moe-router-dtype fp32
-   --moe-grouped-gemm
-)
diff --git a/scripts/models/inkling-small.py b/scripts/models/inkling-small.py
new file mode 100644
index 00000000000..e1d0226515b
--- /dev/null
+++ b/scripts/models/inkling-small.py
@@ -0,0 +1,44 @@
+import os
+
+
+def model_args(nlayers: int | None = None) -> str:
+    # Inkling-Small 276B config (42 layers; derived from the HF config the same way
+    # inkling.py maps the Inkling one).
+    nlayers = nlayers if nlayers is not None else int(os.environ.get("MODEL_ARGS_NUM_LAYERS") or 42)
+    return (
+        "--disable-bias-linear "
+        f"--num-layers {nlayers} "
+        "--hidden-size 4096 "
+        "--ffn-hidden-size 2048 "
+        "--num-attention-heads 32 "
+        "--group-query-attention "
+        "--num-query-groups 8 "
+        "--kv-channels 128 "
+        "--normalization RMSNorm "
+        "--norm-epsilon 1e-6 "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--vocab-size 201024 "
+        "--hidden-dropout 0.0 "
+        "--attention-dropout 0.0 "
+        "--attention-softmax-in-fp32 "
+        "--position-embedding-type none "
+        "--no-rope-fusion "
+        "--no-masked-softmax-fusion "
+        "--max-position-embeddings 1048576 "
+        # MoE
+        "--num-experts 256 "
+        "--moe-ffn-hidden-size 2048 "
+        "--moe-router-topk 6 "
+        "--moe-shared-expert-intermediate-size 2048 "
+        "--moe-router-pre-softmax "
+        "--moe-router-score-function sigmoid "
+        "--moe-router-enable-expert-bias "
+        "--moe-router-load-balancing-type seq_aux_loss "
+        "--moe-token-dispatcher-type alltoall "
+        "--moe-aux-loss-coeff 0 "
+        "--moe-grouped-gemm "
+        "--qk-layernorm "
+        # Inkling model provider
+        "--custom-model-provider-path miles_plugins.models.inkling.model.inkling_model_provider "
+    )
diff --git a/scripts/models/inkling-small.sh b/scripts/models/inkling-small.sh
deleted file mode 100644
index 10131973de1..00000000000
--- a/scripts/models/inkling-small.sh
+++ /dev/null
@@ -1,43 +0,0 @@
-NLAYERS="${MODEL_ARGS_NUM_LAYERS:-42}"
-
-# Inkling-Small 276B config (42 layers; derived from the HF config the same way
-# inkling.sh maps the Inkling one).
-MODEL_ARGS=(
-    --disable-bias-linear
-    --num-layers $NLAYERS
-    --hidden-size 4096
-    --ffn-hidden-size 2048
-    --num-attention-heads 32
-    --group-query-attention
-    --num-query-groups 8
-    --kv-channels 128
-    --normalization RMSNorm
-    --norm-epsilon 1e-6
-    --swiglu
-    --untie-embeddings-and-output-weights
-    --vocab-size 201024
-    --hidden-dropout 0.0
-    --attention-dropout 0.0
-    --attention-softmax-in-fp32
-    --position-embedding-type none
-    --no-rope-fusion
-    --no-masked-softmax-fusion
-    --max-position-embeddings 1048576
-
-    # MoE
-    --num-experts 256
-    --moe-ffn-hidden-size 2048
-    --moe-router-topk 6
-    --moe-shared-expert-intermediate-size 2048
-    --moe-router-pre-softmax
-    --moe-router-score-function sigmoid
-    --moe-router-enable-expert-bias
-    --moe-router-load-balancing-type seq_aux_loss
-    --moe-token-dispatcher-type alltoall
-    --moe-aux-loss-coeff 0
-    --moe-grouped-gemm
-    --qk-layernorm
-
-    # Inkling model provider
-    --custom-model-provider-path miles_plugins.models.inkling.model.inkling_model_provider
-)
diff --git a/scripts/models/inkling.py b/scripts/models/inkling.py
new file mode 100644
index 00000000000..d5b2e726287
--- /dev/null
+++ b/scripts/models/inkling.py
@@ -0,0 +1,43 @@
+import os
+
+
+def model_args(nlayers: int | None = None) -> str:
+    # Inkling config (66L full model; set MODEL_ARGS_NUM_LAYERS=4 for the 4-layer slice)
+    nlayers = nlayers if nlayers is not None else int(os.environ.get("MODEL_ARGS_NUM_LAYERS") or 66)
+    return (
+        "--disable-bias-linear "
+        f"--num-layers {nlayers} "
+        "--hidden-size 6144 "
+        "--ffn-hidden-size 3072 "
+        "--num-attention-heads 64 "
+        "--group-query-attention "
+        "--num-query-groups 8 "
+        "--kv-channels 128 "
+        "--normalization RMSNorm "
+        "--norm-epsilon 1e-6 "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--vocab-size 201024 "
+        "--hidden-dropout 0.0 "
+        "--attention-dropout 0.0 "
+        "--attention-softmax-in-fp32 "
+        "--position-embedding-type none "
+        "--no-rope-fusion "
+        "--no-masked-softmax-fusion "
+        "--max-position-embeddings 1048576 "
+        # MoE
+        "--num-experts 256 "
+        "--moe-ffn-hidden-size 3072 "
+        "--moe-router-topk 6 "
+        "--moe-shared-expert-intermediate-size 3072 "
+        "--moe-router-pre-softmax "
+        "--moe-router-score-function sigmoid "
+        "--moe-router-enable-expert-bias "
+        "--moe-router-load-balancing-type seq_aux_loss "
+        "--moe-token-dispatcher-type alltoall "
+        "--moe-aux-loss-coeff 0 "
+        "--moe-grouped-gemm "
+        "--qk-layernorm "
+        # Inkling model provider
+        "--custom-model-provider-path miles_plugins.models.inkling.model.inkling_model_provider "
+    )
diff --git a/scripts/models/inkling.sh b/scripts/models/inkling.sh
deleted file mode 100644
index 313f4a3db88..00000000000
--- a/scripts/models/inkling.sh
+++ /dev/null
@@ -1,42 +0,0 @@
-NLAYERS="${MODEL_ARGS_NUM_LAYERS:-66}"
-
-# Inkling config (66L full model; set MODEL_ARGS_NUM_LAYERS=4 for the 4-layer slice)
-MODEL_ARGS=(
-    --disable-bias-linear
-    --num-layers $NLAYERS
-    --hidden-size 6144
-    --ffn-hidden-size 3072
-    --num-attention-heads 64
-    --group-query-attention
-    --num-query-groups 8
-    --kv-channels 128
-    --normalization RMSNorm
-    --norm-epsilon 1e-6
-    --swiglu
-    --untie-embeddings-and-output-weights
-    --vocab-size 201024
-    --hidden-dropout 0.0
-    --attention-dropout 0.0
-    --attention-softmax-in-fp32
-    --position-embedding-type none
-    --no-rope-fusion
-    --no-masked-softmax-fusion
-    --max-position-embeddings 1048576
-
-    # MoE
-    --num-experts 256
-    --moe-ffn-hidden-size 3072
-    --moe-router-topk 6
-    --moe-shared-expert-intermediate-size 3072
-    --moe-router-pre-softmax
-    --moe-router-score-function sigmoid
-    --moe-router-enable-expert-bias
-    --moe-router-load-balancing-type seq_aux_loss
-    --moe-token-dispatcher-type alltoall
-    --moe-aux-loss-coeff 0
-    --moe-grouped-gemm
-    --qk-layernorm
-
-    # Inkling model provider
-    --custom-model-provider-path miles_plugins.models.inkling.model.inkling_model_provider
-)
diff --git a/scripts/models/joyai-llm-flash.py b/scripts/models/joyai-llm-flash.py
new file mode 100644
index 00000000000..14296e4502c
--- /dev/null
+++ b/scripts/models/joyai-llm-flash.py
@@ -0,0 +1,55 @@
+import os
+
+from model_args_utils import moe_layer_freq
+
+
+FIRST_K_DENSE_REPLACE = 1
+
+
+def model_args(nlayers: int | None = None, rotary_base: str | None = None) -> str:
+    nlayers = nlayers if nlayers is not None else int(os.environ.get("MODEL_ARGS_NUM_LAYERS") or 40)
+    rotary_base = rotary_base if rotary_base is not None else os.environ.get("MODEL_ARGS_ROTARY_BASE") or "32000000"
+    return (
+        "--disable-bias-linear "
+        f"--num-layers {nlayers} "
+        "--hidden-size 2048 "
+        "--ffn-hidden-size 7168 "
+        "--num-attention-heads 32 "
+        "--kv-channels 128 "
+        "--normalization RMSNorm "
+        "--position-embedding-type rope "
+        "--norm-epsilon 1e-6 "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--vocab-size 129280 "
+        "--multi-latent-attention "
+        "--q-lora-rank 1536 "
+        "--kv-lora-rank 512 "
+        "--qk-head-dim 128 "
+        "--qk-pos-emb-head-dim 64 "
+        "--v-head-dim 128 "
+        "--qk-layernorm "
+        f"--rotary-base {rotary_base} "
+        "--mscale 1.0 "
+        "--mscale-all-dim 1.0 "
+        "--attention-softmax-in-fp32 "
+        "--no-rope-fusion "
+        "--num-experts 256 "
+        f"--moe-layer-freq {moe_layer_freq(nlayers=nlayers, first_k_dense_replace=FIRST_K_DENSE_REPLACE)} "
+        "--moe-ffn-hidden-size 768 "
+        "--moe-router-topk 8 "
+        "--moe-shared-expert-intermediate-size 768 "
+        "--moe-router-pre-softmax "
+        "--moe-router-score-function sigmoid "
+        "--moe-router-enable-expert-bias "
+        "--moe-router-load-balancing-type seq_aux_loss "
+        "--moe-token-dispatcher-type alltoall "
+        "--moe-aux-loss-coeff 0 "
+        "--moe-router-bias-update-rate 0 "
+        "--moe-router-group-topk 1 "
+        "--moe-router-num-groups 1 "
+        "--moe-grouped-gemm "
+        "--moe-router-topk-scaling-factor 2.5 "
+        "--moe-router-dtype fp32 "
+        "--moe-permute-fusion "
+    )
diff --git a/scripts/models/joyai-llm-flash.sh b/scripts/models/joyai-llm-flash.sh
deleted file mode 100644
index 8a2b28c537a..00000000000
--- a/scripts/models/joyai-llm-flash.sh
+++ /dev/null
@@ -1,60 +0,0 @@
-NLAYERS="${MODEL_ARGS_NUM_LAYERS:-40}"
-FIRST_K_DENSE_REPLACE=1
-
-arr=()
-for ((i=0; i<NLAYERS; i++)); do
-  if (( i < FIRST_K_DENSE_REPLACE )); then
-    arr+=(0)
-  else
-    arr+=(1)
-  fi
-done
-
-printf -v MOE_LAYER_FREQ "[%s]" "$(IFS=', '; echo "${arr[*]}")"
-
-MODEL_ARGS=(
-    --disable-bias-linear
-    --num-layers $NLAYERS
-    --hidden-size 2048
-    --ffn-hidden-size 7168
-    --num-attention-heads 32
-    --kv-channels 128
-    --normalization RMSNorm
-    --position-embedding-type rope
-    --norm-epsilon 1e-6
-    --swiglu
-    --untie-embeddings-and-output-weights
-    --vocab-size 129280
-
-    --multi-latent-attention
-    --q-lora-rank 1536
-    --kv-lora-rank 512
-    --qk-head-dim 128
-    --qk-pos-emb-head-dim 64
-    --v-head-dim 128
-    --qk-layernorm
-    --rotary-base "${MODEL_ARGS_ROTARY_BASE:-32000000}"
-    --mscale 1.0
-    --mscale-all-dim 1.0
-    --attention-softmax-in-fp32
-    --no-rope-fusion
-
-    --num-experts 256
-    --moe-layer-freq $MOE_LAYER_FREQ
-    --moe-ffn-hidden-size 768
-    --moe-router-topk 8
-    --moe-shared-expert-intermediate-size 768
-    --moe-router-pre-softmax
-    --moe-router-score-function sigmoid
-    --moe-router-enable-expert-bias
-    --moe-router-load-balancing-type seq_aux_loss
-    --moe-token-dispatcher-type alltoall
-    --moe-aux-loss-coeff 0
-    --moe-router-bias-update-rate 0
-    --moe-router-group-topk 1
-    --moe-router-num-groups 1
-    --moe-grouped-gemm
-    --moe-router-topk-scaling-factor 2.5
-    --moe-router-dtype fp32
-    --moe-permute-fusion
-)
diff --git a/scripts/models/kimi-k2-thinking.py b/scripts/models/kimi-k2-thinking.py
new file mode 100644
index 00000000000..a5e00b7a1e6
--- /dev/null
+++ b/scripts/models/kimi-k2-thinking.py
@@ -0,0 +1,50 @@
+from model_args_utils import moe_layer_freq
+
+
+def model_args(nlayers: int = 61, first_k_dense_replace: int = 1) -> str:
+    return (
+        "--disable-bias-linear "
+        f"--num-layers {nlayers} "
+        "--hidden-size 7168 "
+        "--ffn-hidden-size 18432 "
+        "--num-attention-heads 64 "
+        "--kv-channels 64 "
+        "--normalization RMSNorm "
+        "--position-embedding-type rope "
+        "--norm-epsilon 1e-5 "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--vocab-size 163840 "
+        "--multi-latent-attention "
+        "--q-lora-rank 1536 "
+        "--kv-lora-rank 512 "
+        "--qk-head-dim 128 "
+        "--qk-pos-emb-head-dim 64 "
+        "--v-head-dim 128 "
+        "--qk-layernorm "
+        "--rotary-scaling-factor 64.0 "
+        "--rotary-base 50000 "
+        "--mscale 1.0 "
+        "--mscale-all-dim 1.0 "
+        "--attention-softmax-in-fp32 "
+        "--no-rope-fusion "
+        # moe
+        "--num-experts 384 "
+        f"--moe-layer-freq {moe_layer_freq(nlayers=nlayers, first_k_dense_replace=first_k_dense_replace)} "
+        "--moe-ffn-hidden-size 2048 "
+        "--moe-router-topk 8 "
+        "--moe-shared-expert-intermediate-size 2048 "
+        "--moe-router-pre-softmax "
+        "--moe-router-score-function sigmoid "
+        "--moe-router-enable-expert-bias "
+        "--moe-router-load-balancing-type seq_aux_loss "
+        "--moe-token-dispatcher-type alltoall "
+        "--moe-aux-loss-coeff 0 "
+        "--moe-router-bias-update-rate 0 "
+        "--moe-router-group-topk 1 "
+        "--moe-router-num-groups 1 "
+        "--moe-grouped-gemm "
+        "--moe-router-topk-scaling-factor 2.827 "
+        "--moe-router-dtype fp32 "
+        "--moe-permute-fusion "
+    )
diff --git a/scripts/models/kimi-k2-thinking.sh b/scripts/models/kimi-k2-thinking.sh
deleted file mode 100644
index b7fceda5913..00000000000
--- a/scripts/models/kimi-k2-thinking.sh
+++ /dev/null
@@ -1,63 +0,0 @@
-NLAYERS=61
-FIRST_K_DENSE_REPLACE=1
-
-arr=()
-for ((i=0; i<NLAYERS; i++)); do
-  if (( i < FIRST_K_DENSE_REPLACE )); then
-    arr+=(0)
-  else
-    arr+=(1)
-  fi
-done
-
-printf -v MOE_LAYER_FREQ "[%s]" "$(IFS=', '; echo "${arr[*]}")"
-
-# kimi-k2-thinking
-MODEL_ARGS=(
-    --disable-bias-linear
-    --num-layers 61
-    --hidden-size 7168
-    --ffn-hidden-size 18432
-    --num-attention-heads 64
-    --kv-channels 64
-    --normalization RMSNorm
-    --position-embedding-type rope
-    --norm-epsilon 1e-5
-    --swiglu
-    --untie-embeddings-and-output-weights
-    --vocab-size 163840
-    
-    --multi-latent-attention
-    --q-lora-rank 1536
-    --kv-lora-rank 512
-    --qk-head-dim 128
-    --qk-pos-emb-head-dim 64
-    --v-head-dim 128
-    --qk-layernorm
-    --rotary-scaling-factor 64.0
-    --rotary-base 50000
-    --mscale 1.0
-    --mscale-all-dim 1.0
-    --attention-softmax-in-fp32
-    --no-rope-fusion
-
-    # moe
-    --num-experts 384
-    --moe-layer-freq "$MOE_LAYER_FREQ"
-    --moe-ffn-hidden-size 2048
-    --moe-router-topk 8
-    --moe-shared-expert-intermediate-size 2048
-    --moe-router-pre-softmax
-    --moe-router-score-function sigmoid
-    --moe-router-enable-expert-bias
-    --moe-router-load-balancing-type seq_aux_loss
-    --moe-token-dispatcher-type alltoall
-    --moe-aux-loss-coeff 0
-    --moe-router-bias-update-rate 0
-    --moe-router-group-topk 1
-    --moe-router-num-groups 1
-    --moe-grouped-gemm
-    --moe-router-topk-scaling-factor 2.827
-    --moe-router-dtype fp32
-    --moe-permute-fusion
-)
\ No newline at end of file
diff --git a/scripts/models/kimi-k2.py b/scripts/models/kimi-k2.py
new file mode 100644
index 00000000000..7100124f290
--- /dev/null
+++ b/scripts/models/kimi-k2.py
@@ -0,0 +1,54 @@
+from model_args_utils import moe_layer_freq
+
+
+NLAYERS = 61
+FIRST_K_DENSE_REPLACE = 1
+
+
+def model_args() -> str:
+    return (
+        "--disable-bias-linear "
+        "--num-layers 61 "
+        "--hidden-size 7168 "
+        "--ffn-hidden-size 18432 "
+        "--num-attention-heads 64 "
+        "--kv-channels 64 "
+        "--normalization RMSNorm "
+        "--position-embedding-type rope "
+        "--norm-epsilon 1e-6 "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--vocab-size 163840 "
+        "--multi-latent-attention "
+        "--q-lora-rank 1536 "
+        "--kv-lora-rank 512 "
+        "--qk-head-dim 128 "
+        "--qk-pos-emb-head-dim 64 "
+        "--v-head-dim 128 "
+        "--qk-layernorm "
+        "--rotary-scaling-factor 32.0 "
+        "--rotary-base 50000 "
+        "--mscale 1.0 "
+        "--mscale-all-dim 1.0 "
+        "--attention-softmax-in-fp32 "
+        "--no-rope-fusion "
+        # moe
+        "--num-experts 384 "
+        f"--moe-layer-freq {moe_layer_freq(nlayers=NLAYERS, first_k_dense_replace=FIRST_K_DENSE_REPLACE)} "
+        "--moe-ffn-hidden-size 2048 "
+        "--moe-router-topk 8 "
+        "--moe-shared-expert-intermediate-size 2048 "
+        "--moe-router-pre-softmax "
+        "--moe-router-score-function sigmoid "
+        "--moe-router-enable-expert-bias "
+        "--moe-router-load-balancing-type seq_aux_loss "
+        "--moe-token-dispatcher-type alltoall "
+        "--moe-aux-loss-coeff 0 "
+        "--moe-router-bias-update-rate 0 "
+        "--moe-router-group-topk 1 "
+        "--moe-router-num-groups 1 "
+        "--moe-grouped-gemm "
+        "--moe-router-topk-scaling-factor 2.827 "
+        "--moe-router-dtype fp32 "
+        "--moe-permute-fusion "
+    )
diff --git a/scripts/models/kimi-k2.sh b/scripts/models/kimi-k2.sh
deleted file mode 100644
index eafb7acadd1..00000000000
--- a/scripts/models/kimi-k2.sh
+++ /dev/null
@@ -1,63 +0,0 @@
-NLAYERS=61
-FIRST_K_DENSE_REPLACE=1
-
-arr=()
-for ((i=0; i<NLAYERS; i++)); do
-  if (( i < FIRST_K_DENSE_REPLACE )); then
-    arr+=(0)
-  else
-    arr+=(1)
-  fi
-done
-
-printf -v MOE_LAYER_FREQ "[%s]" "$(IFS=', '; echo "${arr[*]}")"
-
-# kimi-k2
-MODEL_ARGS=(
-    --disable-bias-linear
-    --num-layers 61
-    --hidden-size 7168
-    --ffn-hidden-size 18432
-    --num-attention-heads 64
-    --kv-channels 64
-    --normalization RMSNorm
-    --position-embedding-type rope
-    --norm-epsilon 1e-6
-    --swiglu
-    --untie-embeddings-and-output-weights
-    --vocab-size 163840
-    
-    --multi-latent-attention
-    --q-lora-rank 1536
-    --kv-lora-rank 512
-    --qk-head-dim 128
-    --qk-pos-emb-head-dim 64
-    --v-head-dim 128
-    --qk-layernorm
-    --rotary-scaling-factor 32.0
-    --rotary-base 50000
-    --mscale 1.0
-    --mscale-all-dim 1.0
-    --attention-softmax-in-fp32
-    --no-rope-fusion
-
-    # moe
-    --num-experts 384
-    --moe-layer-freq "$MOE_LAYER_FREQ"
-    --moe-ffn-hidden-size 2048
-    --moe-router-topk 8
-    --moe-shared-expert-intermediate-size 2048
-    --moe-router-pre-softmax
-    --moe-router-score-function sigmoid
-    --moe-router-enable-expert-bias
-    --moe-router-load-balancing-type seq_aux_loss
-    --moe-token-dispatcher-type alltoall
-    --moe-aux-loss-coeff 0
-    --moe-router-bias-update-rate 0
-    --moe-router-group-topk 1
-    --moe-router-num-groups 1
-    --moe-grouped-gemm
-    --moe-router-topk-scaling-factor 2.827
-    --moe-router-dtype fp32
-    --moe-permute-fusion
-)
\ No newline at end of file
diff --git a/scripts/models/kimi-k25_2layer.py b/scripts/models/kimi-k25_2layer.py
new file mode 100644
index 00000000000..a98cbbabb6f
--- /dev/null
+++ b/scripts/models/kimi-k25_2layer.py
@@ -0,0 +1,9 @@
+from model_args_utils import load_sibling_model_args
+
+
+def model_args() -> str:
+    # Override for the 2-layer pruned debugging model (first_k_dense_replace=1):
+    # 1 dense layer + 1 MoE layer. Architecture is otherwise identical to the full
+    # Kimi-K2.5 / K2-Thinking, so we reuse those MODEL_ARGS and only patch the
+    # layer count and the MoE-layer-frequency mask.
+    return load_sibling_model_args(__file__, "kimi-k2-thinking", nlayers=2, first_k_dense_replace=1)
diff --git a/scripts/models/kimi-k25_2layer.sh b/scripts/models/kimi-k25_2layer.sh
deleted file mode 100644
index f57c207b3a8..00000000000
--- a/scripts/models/kimi-k25_2layer.sh
+++ /dev/null
@@ -1,26 +0,0 @@
-SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]:-$0}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/kimi-k2-thinking.sh"
-
-# Override for the 2-layer pruned debugging model (first_k_dense_replace=1):
-# 1 dense layer + 1 MoE layer. Architecture is otherwise identical to the full
-# Kimi-K2.5 / K2-Thinking, so we reuse those MODEL_ARGS and only patch the
-# layer count and the MoE-layer-frequency mask.
-NLAYERS=2
-FIRST_K_DENSE_REPLACE=1
-
-arr=()
-for ((i = 0; i < NLAYERS; i++)); do
-    if ((i < FIRST_K_DENSE_REPLACE)); then
-        arr+=(0)
-    else
-        arr+=(1)
-    fi
-done
-printf -v MOE_LAYER_FREQ "[%s]" "$(IFS=', '; echo "${arr[*]}")"
-
-for ((i = 0; i < ${#MODEL_ARGS[@]}; i++)); do
-    case "${MODEL_ARGS[$i]}" in
-        --num-layers) MODEL_ARGS[$((i + 1))]=$NLAYERS ;;
-        --moe-layer-freq) MODEL_ARGS[$((i + 1))]="$MOE_LAYER_FREQ" ;;
-    esac
-done
diff --git a/scripts/models/llama3.1-8B-Instruct.py b/scripts/models/llama3.1-8B-Instruct.py
new file mode 100644
index 00000000000..1375255d845
--- /dev/null
+++ b/scripts/models/llama3.1-8B-Instruct.py
@@ -0,0 +1,21 @@
+def model_args() -> str:
+    return (
+        "--swiglu "
+        "--num-layers 32 "
+        "--hidden-size 4096 "
+        "--ffn-hidden-size 14336 "
+        "--num-attention-heads 32 "
+        "--group-query-attention "
+        "--num-query-groups 8 "
+        "--max-position-embeddings 131072 "
+        "--use-rotary-position-embeddings "
+        "--disable-bias-linear "
+        "--normalization RMSNorm "
+        "--norm-epsilon 1e-5 "
+        "--rotary-base 500000 "
+        "--vocab-size 128256 "
+        "--kv-channels 128 "
+        "--use-rope-scaling "
+        "--rotary-scaling-factor 8.0 "
+        "--untie-embeddings-and-output-weights "
+    )
diff --git a/scripts/models/llama3.1-8B-Instruct.sh b/scripts/models/llama3.1-8B-Instruct.sh
deleted file mode 100644
index 0815b3e0a34..00000000000
--- a/scripts/models/llama3.1-8B-Instruct.sh
+++ /dev/null
@@ -1,20 +0,0 @@
-MODEL_ARGS=(
-   --swiglu
-   --num-layers 32
-   --hidden-size 4096
-   --ffn-hidden-size 14336
-   --num-attention-heads 32
-   --group-query-attention
-   --num-query-groups 8
-   --max-position-embeddings 131072
-   --use-rotary-position-embeddings
-   --disable-bias-linear
-   --normalization "RMSNorm"
-   --norm-epsilon 1e-5
-   --rotary-base 500000
-   --vocab-size 128256
-   --kv-channels 128
-   --use-rope-scaling
-   --rotary-scaling-factor 8.0
-   --untie-embeddings-and-output-weights
-)
\ No newline at end of file
diff --git a/scripts/models/llama3.2-3B-Instruct-amd.py b/scripts/models/llama3.2-3B-Instruct-amd.py
new file mode 100644
index 00000000000..ba185446bee
--- /dev/null
+++ b/scripts/models/llama3.2-3B-Instruct-amd.py
@@ -0,0 +1,20 @@
+def model_args() -> str:
+    return (
+        "--swiglu "
+        "--num-layers 28 "
+        "--hidden-size 3072 "
+        "--ffn-hidden-size 8192 "
+        "--num-attention-heads 24 "
+        "--group-query-attention "
+        "--num-query-groups 8 "
+        "--max-position-embeddings 131072 "
+        "--use-rotary-position-embeddings "
+        "--disable-bias-linear "
+        "--normalization RMSNorm "
+        "--norm-epsilon 1e-5 "
+        "--rotary-base 500000 "
+        "--vocab-size 128256 "
+        "--kv-channels 128 "
+        "--use-rope-scaling "
+        "--rotary-scaling-factor 32.0 "
+    )
diff --git a/scripts/models/llama3.2-3B-Instruct-amd.sh b/scripts/models/llama3.2-3B-Instruct-amd.sh
deleted file mode 100644
index 654de5a3865..00000000000
--- a/scripts/models/llama3.2-3B-Instruct-amd.sh
+++ /dev/null
@@ -1,19 +0,0 @@
-MODEL_ARGS=(
-   --swiglu
-   --num-layers 28
-   --hidden-size 3072
-   --ffn-hidden-size 8192
-   --num-attention-heads 24
-   --group-query-attention
-   --num-query-groups 8
-   --max-position-embeddings 131072
-   --use-rotary-position-embeddings
-   --disable-bias-linear
-   --normalization "RMSNorm"
-   --norm-epsilon 1e-5
-   --rotary-base 500000
-   --vocab-size 128256
-   --kv-channels 128
-   --use-rope-scaling
-   --rotary-scaling-factor 32.0
-)
diff --git a/scripts/models/llama3.2-3B-Instruct.py b/scripts/models/llama3.2-3B-Instruct.py
new file mode 100644
index 00000000000..ba185446bee
--- /dev/null
+++ b/scripts/models/llama3.2-3B-Instruct.py
@@ -0,0 +1,20 @@
+def model_args() -> str:
+    return (
+        "--swiglu "
+        "--num-layers 28 "
+        "--hidden-size 3072 "
+        "--ffn-hidden-size 8192 "
+        "--num-attention-heads 24 "
+        "--group-query-attention "
+        "--num-query-groups 8 "
+        "--max-position-embeddings 131072 "
+        "--use-rotary-position-embeddings "
+        "--disable-bias-linear "
+        "--normalization RMSNorm "
+        "--norm-epsilon 1e-5 "
+        "--rotary-base 500000 "
+        "--vocab-size 128256 "
+        "--kv-channels 128 "
+        "--use-rope-scaling "
+        "--rotary-scaling-factor 32.0 "
+    )
diff --git a/scripts/models/llama3.2-3B-Instruct.sh b/scripts/models/llama3.2-3B-Instruct.sh
deleted file mode 100644
index ff50130c8a7..00000000000
--- a/scripts/models/llama3.2-3B-Instruct.sh
+++ /dev/null
@@ -1,19 +0,0 @@
-MODEL_ARGS=(
-   --swiglu
-   --num-layers 28
-   --hidden-size 3072
-   --ffn-hidden-size 8192
-   --num-attention-heads 24
-   --group-query-attention
-   --num-query-groups 8
-   --max-position-embeddings 131072
-   --use-rotary-position-embeddings
-   --disable-bias-linear
-   --normalization "RMSNorm"
-   --norm-epsilon 1e-5
-   --rotary-base 500000
-   --vocab-size 128256
-   --kv-channels 128
-   --use-rope-scaling
-   --rotary-scaling-factor 32.0
-)
\ No newline at end of file
diff --git a/scripts/models/mimo-7B-rl.py b/scripts/models/mimo-7B-rl.py
new file mode 100644
index 00000000000..d790dbddbc7
--- /dev/null
+++ b/scripts/models/mimo-7B-rl.py
@@ -0,0 +1,20 @@
+def model_args() -> str:
+    return (
+        "--swiglu "
+        "--num-layers 36 "
+        "--hidden-size 4096 "
+        "--ffn-hidden-size 11008 "
+        "--num-attention-heads 32 "
+        "--group-query-attention "
+        "--num-query-groups 8 "
+        "--use-rotary-position-embeddings "
+        "--disable-bias-linear "
+        "--add-qkv-bias "
+        "--normalization RMSNorm "
+        "--norm-epsilon 1e-05 "
+        "--rotary-base 640000 "
+        "--vocab-size 151680 "
+        "--untie-embeddings-and-output-weights "
+        "--max-position-embeddings 32768 "
+        "--mtp-num-layers 1 "
+    )
diff --git a/scripts/models/mimo-7B-rl.sh b/scripts/models/mimo-7B-rl.sh
deleted file mode 100644
index 22366935f99..00000000000
--- a/scripts/models/mimo-7B-rl.sh
+++ /dev/null
@@ -1,19 +0,0 @@
-MODEL_ARGS=(
-    --swiglu
-    --num-layers 36
-    --hidden-size 4096
-    --ffn-hidden-size 11008
-    --num-attention-heads 32
-    --group-query-attention
-    --num-query-groups 8
-    --use-rotary-position-embeddings
-    --disable-bias-linear
-    --add-qkv-bias
-    --normalization "RMSNorm"
-    --norm-epsilon 1e-05
-    --rotary-base 640000
-    --vocab-size 151680
-    --untie-embeddings-and-output-weights
-    --max-position-embeddings 32768
-    --mtp-num-layers 1
-)
diff --git a/scripts/models/moonlight.py b/scripts/models/moonlight.py
new file mode 100644
index 00000000000..432ef0b6a16
--- /dev/null
+++ b/scripts/models/moonlight.py
@@ -0,0 +1,60 @@
+from model_args_utils import moe_layer_freq
+
+
+MOE_SHARED_EXPERTS = 2
+MOE_FFN_HIDDEN = 1408
+MOE_SHARED_EXPERT_INTERMEDIATE_SIZE = MOE_FFN_HIDDEN * MOE_SHARED_EXPERTS
+MOE_ROUTER_TOPK_SCALING_FACTOR = 2.446
+NLAYERS = 27
+FIRST_K_DENSE_REPLACE = 1
+
+
+def model_args() -> str:
+    return (
+        "--disable-bias-linear "
+        "--num-layers 27 "
+        "--hidden-size 2048 "
+        "--ffn-hidden-size 11264 "
+        "--num-attention-heads 16 "
+        "--kv-channels 128 "
+        "--normalization RMSNorm "
+        "--position-embedding-type rope "
+        "--norm-epsilon 1e-5 "
+        "--rotary-percent 1.0 "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--no-masked-softmax-fusion "
+        "--vocab-size 163840 "
+        "--multi-latent-attention "
+        "--kv-lora-rank 512 "
+        "--qk-head-dim 128 "
+        "--qk-pos-emb-head-dim 64 "
+        "--v-head-dim 128 "
+        "--qk-layernorm "
+        "--rotary-scaling-factor 1 "
+        "--rotary-base 50000 "
+        "--mscale 1.0 "
+        "--mscale-all-dim 1.0 "
+        "--attention-softmax-in-fp32 "
+        "--no-rope-fusion "
+        # moe
+        "--num-experts 64 "
+        f"--moe-layer-freq {moe_layer_freq(nlayers=NLAYERS, first_k_dense_replace=FIRST_K_DENSE_REPLACE)} "
+        f"--moe-ffn-hidden-size {MOE_FFN_HIDDEN} "
+        "--moe-router-topk 6 "
+        f"--moe-shared-expert-intermediate-size {MOE_SHARED_EXPERT_INTERMEDIATE_SIZE} "
+        "--moe-router-pre-softmax "
+        "--moe-router-score-function sigmoid "
+        "--moe-router-enable-expert-bias "
+        "--moe-router-load-balancing-type seq_aux_loss "
+        "--moe-token-dispatcher-type alltoall "
+        "--moe-aux-loss-coeff 0 "
+        "--moe-router-bias-update-rate 0 "
+        "--moe-router-group-topk 1 "
+        "--moe-router-num-groups 1 "
+        "--moe-grouped-gemm "
+        f"--moe-router-topk-scaling-factor {MOE_ROUTER_TOPK_SCALING_FACTOR} "
+        "--moe-token-drop-policy probs "
+        "--moe-router-dtype fp32 "
+        "--moe-permute-fusion "
+    )
diff --git a/scripts/models/moonlight.sh b/scripts/models/moonlight.sh
deleted file mode 100644
index bcce99892a7..00000000000
--- a/scripts/models/moonlight.sh
+++ /dev/null
@@ -1,69 +0,0 @@
-MOE_SHARED_EXPERTS=2
-MOE_FFN_HIDDEN=1408
-MOE_SHARED_EXPERT_INTERMEDIATE_SIZE=$(($MOE_FFN_HIDDEN * $MOE_SHARED_EXPERTS))
-MOE_ROUTER_TOPK_SCALING_FACTOR=2.446
-NLAYERS=27
-FIRST_K_DENSE_REPLACE=1
-
-arr=()
-for ((i=0; i<NLAYERS; i++)); do
-  if (( i < FIRST_K_DENSE_REPLACE )); then
-    arr+=(0)
-  else
-    arr+=(1)
-  fi
-done
-
-printf -v MOE_LAYER_FREQ "[%s]" "$(IFS=', '; echo "${arr[*]}")"
-
-# moonlight
-MODEL_ARGS=(
-    --disable-bias-linear
-    --num-layers 27
-    --hidden-size 2048
-    --ffn-hidden-size 11264
-    --num-attention-heads 16
-    --kv-channels 128
-    --normalization RMSNorm
-    --position-embedding-type rope
-    --norm-epsilon 1e-5
-    --rotary-percent 1.0
-    --swiglu
-    --untie-embeddings-and-output-weights
-    --no-masked-softmax-fusion
-    --vocab-size 163840
-
-    --multi-latent-attention
-    --kv-lora-rank 512
-    --qk-head-dim 128
-    --qk-pos-emb-head-dim 64
-    --v-head-dim 128
-    --qk-layernorm
-    --rotary-scaling-factor 1
-    --rotary-base 50000
-    --mscale 1.0
-    --mscale-all-dim 1.0
-    --attention-softmax-in-fp32
-    --no-rope-fusion
-
-    # moe
-    --num-experts 64
-    --moe-layer-freq "$MOE_LAYER_FREQ"
-    --moe-ffn-hidden-size $MOE_FFN_HIDDEN
-    --moe-router-topk 6
-    --moe-shared-expert-intermediate-size $MOE_SHARED_EXPERT_INTERMEDIATE_SIZE
-    --moe-router-pre-softmax
-    --moe-router-score-function sigmoid
-    --moe-router-enable-expert-bias
-    --moe-router-load-balancing-type seq_aux_loss
-    --moe-token-dispatcher-type alltoall
-    --moe-aux-loss-coeff 0
-    --moe-router-bias-update-rate 0
-    --moe-router-group-topk 1
-    --moe-router-num-groups 1
-    --moe-grouped-gemm
-    --moe-router-topk-scaling-factor $MOE_ROUTER_TOPK_SCALING_FACTOR
-    --moe-token-drop-policy probs
-    --moe-router-dtype fp32
-    --moe-permute-fusion
-)
\ No newline at end of file
diff --git a/scripts/models/nemotron-3-nano-30b-a3b.py b/scripts/models/nemotron-3-nano-30b-a3b.py
new file mode 100644
index 00000000000..8039ec8bdf5
--- /dev/null
+++ b/scripts/models/nemotron-3-nano-30b-a3b.py
@@ -0,0 +1,36 @@
+def model_args() -> str:
+    return (
+        "--disable-bias-linear "
+        "--group-query-attention "
+        "--num-attention-heads 32 "
+        "--num-query-groups 2 "
+        "--kv-channels 128 "
+        "--num-layers 52 "
+        "--hidden-size 2688 "
+        "--ffn-hidden-size 1856 "
+        "--normalization RMSNorm "
+        "--position-embedding-type none "
+        "--vocab-size 131072 "
+        "--make-vocab-size-divisible-by 128 "
+        "--untie-embeddings-and-output-weights "
+        # MoE specifics
+        "--num-experts 128 "
+        "--moe-router-topk 6 "
+        "--moe-ffn-hidden-size 1856 "
+        "--moe-shared-expert-intermediate-size 3712 "
+        "--moe-router-score-function sigmoid "
+        "--moe-router-enable-expert-bias "
+        "--moe-grouped-gemm "
+        "--moe-router-dtype fp32 "
+        # Routing: config has n_group=1 (MoE groups), topk_group=1,
+        # routed_scaling_factor=2.5. `n_groups=8` is Mamba groups — unrelated to MoE.
+        # With n_group=1, group-limited routing is a no-op (single group of 128).
+        "--moe-router-num-groups 1 "
+        "--moe-router-group-topk 1 "
+        "--moe-router-topk-scaling-factor 2.5 "
+        "--moe-router-pre-softmax "
+        # Match glm4.7-flash (known-working MoE RL) settings more closely.
+        "--moe-router-load-balancing-type seq_aux_loss "
+        "--moe-router-bias-update-rate 0 "
+        "--moe-aux-loss-coeff 0 "
+    )
diff --git a/scripts/models/nemotron-3-nano-30b-a3b.sh b/scripts/models/nemotron-3-nano-30b-a3b.sh
deleted file mode 100644
index bf10a50b7cc..00000000000
--- a/scripts/models/nemotron-3-nano-30b-a3b.sh
+++ /dev/null
@@ -1,47 +0,0 @@
-# NVIDIA Nemotron-3-Nano-30B-A3B (BF16, MoE nemotron_h = hybrid Mamba + Attention + MoE).
-# HF config (verified 2026-04-21):
-#   num_hidden_layers=52  hidden_size=2688  num_attention_heads=32  num_key_value_heads=2
-#   head_dim=128  intermediate_size=1856  moe_intermediate_size=1856
-#   n_routed_experts=128  num_experts_per_tok=6  n_shared_experts=1
-#   moe_shared_expert_intermediate_size=3712  sigmoid routing + aux-free expert bias
-# The AutoBridge path (--megatron-to-hf-mode bridge) + miles NemotronHBridge MoE shim
-# (see miles/backends/megatron_utils/__init__.py) construct the provider and
-# HF↔Megatron mapping_registry at load time. Attention-side structural args go
-# in MODEL_ARGS for miles' arg parser.
-
-MODEL_ARGS=(
-   --disable-bias-linear
-   --group-query-attention
-   --num-attention-heads 32
-   --num-query-groups 2
-   --kv-channels 128
-   --num-layers 52
-   --hidden-size 2688
-   --ffn-hidden-size 1856
-   --normalization RMSNorm
-   --position-embedding-type none
-   --vocab-size 131072
-   --make-vocab-size-divisible-by 128
-   --untie-embeddings-and-output-weights
-
-   # MoE specifics
-   --num-experts 128
-   --moe-router-topk 6
-   --moe-ffn-hidden-size 1856
-   --moe-shared-expert-intermediate-size 3712
-   --moe-router-score-function sigmoid
-   --moe-router-enable-expert-bias
-   --moe-grouped-gemm
-   --moe-router-dtype fp32
-   # Routing: config has n_group=1 (MoE groups), topk_group=1,
-   # routed_scaling_factor=2.5. `n_groups=8` is Mamba groups — unrelated to MoE.
-   # With n_group=1, group-limited routing is a no-op (single group of 128).
-   --moe-router-num-groups 1
-   --moe-router-group-topk 1
-   --moe-router-topk-scaling-factor 2.5
-   --moe-router-pre-softmax
-   # Match glm4.7-flash (known-working MoE RL) settings more closely.
-   --moe-router-load-balancing-type seq_aux_loss
-   --moe-router-bias-update-rate 0
-   --moe-aux-loss-coeff 0
-)
diff --git a/scripts/models/nemotron-3-nano-4b.py b/scripts/models/nemotron-3-nano-4b.py
new file mode 100644
index 00000000000..2a91c4d7f6f
--- /dev/null
+++ b/scripts/models/nemotron-3-nano-4b.py
@@ -0,0 +1,16 @@
+def model_args() -> str:
+    return (
+        "--disable-bias-linear "
+        "--group-query-attention "
+        "--num-attention-heads 40 "
+        "--num-query-groups 8 "
+        "--kv-channels 128 "
+        "--num-layers 42 "
+        "--hidden-size 3136 "
+        "--ffn-hidden-size 12544 "
+        "--normalization RMSNorm "
+        "--position-embedding-type none "
+        "--vocab-size 131072 "
+        "--make-vocab-size-divisible-by 128 "
+        "--untie-embeddings-and-output-weights "
+    )
diff --git a/scripts/models/nemotron-3-nano-4b.sh b/scripts/models/nemotron-3-nano-4b.sh
deleted file mode 100644
index 9e1a76b5e1d..00000000000
--- a/scripts/models/nemotron-3-nano-4b.sh
+++ /dev/null
@@ -1,24 +0,0 @@
-# NVIDIA Nemotron-3-Nano-4B (BF16, dense `nemotron_h` = hybrid Mamba + Attention).
-# HF config (verified 2026-04-21):
-#   num_hidden_layers=42  hidden_size=3136  num_attention_heads=40  num_key_value_heads=8
-#   vocab_size=131072     max_position_embeddings=262144   no RoPE   squared-relu FFN
-# The AutoBridge path (--megatron-to-hf-mode bridge) constructs the full Megatron
-# provider from the HF config.json at load time, including all Mamba-specific
-# fields (mamba_num_heads, mamba_state_dim, hybrid_override_pattern, etc.), so we
-# only keep the attention-side structural args here for miles' arg parser.
-
-MODEL_ARGS=(
-   --disable-bias-linear
-   --group-query-attention
-   --num-attention-heads 40
-   --num-query-groups 8
-   --kv-channels 128
-   --num-layers 42
-   --hidden-size 3136
-   --ffn-hidden-size 12544
-   --normalization RMSNorm
-   --position-embedding-type none
-   --vocab-size 131072
-   --make-vocab-size-divisible-by 128
-   --untie-embeddings-and-output-weights
-)
diff --git a/scripts/models/nemotron-3-super-120b-a12b.py b/scripts/models/nemotron-3-super-120b-a12b.py
new file mode 100644
index 00000000000..83dc71b7e7f
--- /dev/null
+++ b/scripts/models/nemotron-3-super-120b-a12b.py
@@ -0,0 +1,42 @@
+def model_args() -> str:
+    return (
+        "--disable-bias-linear "
+        "--group-query-attention "
+        "--num-attention-heads 32 "
+        "--num-query-groups 2 "
+        "--kv-channels 128 "
+        "--num-layers 88 "
+        "--hidden-size 4096 "
+        "--ffn-hidden-size 2688 "
+        "--normalization RMSNorm "
+        "--position-embedding-type none "
+        "--vocab-size 131072 "
+        "--make-vocab-size-divisible-by 128 "
+        "--untie-embeddings-and-output-weights "
+        # MoE specifics
+        "--num-experts 512 "
+        "--moe-router-topk 22 "
+        "--moe-ffn-hidden-size 2688 "
+        "--moe-shared-expert-intermediate-size 5376 "
+        # Super-120B bottlenecks expert input/output through a 1024-dim latent.
+        # Routed experts run on moe_latent_size, NOT hidden_size, with two extra
+        # fc1/fc2 latent projections per MoE layer. The miles NemotronH bridge
+        # surfaces this from HF config; the CLI arg keeps Megatron's parser happy.
+        "--moe-latent-size 1024 "
+        "--moe-router-score-function sigmoid "
+        "--moe-router-enable-expert-bias "
+        "--moe-grouped-gemm "
+        "--moe-router-dtype fp32 "
+        # Routing: HF config has n_group=1 (MoE groups), topk_group=1,
+        # routed_scaling_factor=5.0. With n_group=1, group-limited routing is a
+        # no-op (single group of 512). `n_groups=8` in HF is Mamba groups —
+        # unrelated to MoE.
+        "--moe-router-num-groups 1 "
+        "--moe-router-group-topk 1 "
+        "--moe-router-topk-scaling-factor 5.0 "
+        "--moe-router-pre-softmax "
+        # Match nano-30b-a3b (known-working MoE RL on nemotron_h) settings.
+        "--moe-router-load-balancing-type seq_aux_loss "
+        "--moe-router-bias-update-rate 0 "
+        "--moe-aux-loss-coeff 0 "
+    )
diff --git a/scripts/models/nemotron-3-super-120b-a12b.sh b/scripts/models/nemotron-3-super-120b-a12b.sh
deleted file mode 100644
index 13e3eceb8bb..00000000000
--- a/scripts/models/nemotron-3-super-120b-a12b.sh
+++ /dev/null
@@ -1,54 +0,0 @@
-# NVIDIA Nemotron-3-Super-120B-A12B (BF16, MoE nemotron_h = hybrid Mamba + Attention + MoE).
-# HF config (verified 2026-05-01):
-#   num_hidden_layers=88  hidden_size=4096  num_attention_heads=32  num_key_value_heads=2
-#   head_dim=128  intermediate_size=2688  moe_intermediate_size=2688
-#   n_routed_experts=512  num_experts_per_tok=22  n_shared_experts=1
-#   moe_shared_expert_intermediate_size=5376  routed_scaling_factor=5.0
-#   n_group=1  topk_group=1  sigmoid routing + aux-free expert bias
-# The AutoBridge path (--megatron-to-hf-mode bridge) + miles NemotronHBridge MoE shim
-# (see miles_plugins/megatron_bridge/nemotron_h.py) construct the provider and
-# HF<->Megatron mapping_registry at load time. Attention-side structural args go
-# in MODEL_ARGS for miles' arg parser.
-
-MODEL_ARGS=(
-   --disable-bias-linear
-   --group-query-attention
-   --num-attention-heads 32
-   --num-query-groups 2
-   --kv-channels 128
-   --num-layers 88
-   --hidden-size 4096
-   --ffn-hidden-size 2688
-   --normalization RMSNorm
-   --position-embedding-type none
-   --vocab-size 131072
-   --make-vocab-size-divisible-by 128
-   --untie-embeddings-and-output-weights
-
-   # MoE specifics
-   --num-experts 512
-   --moe-router-topk 22
-   --moe-ffn-hidden-size 2688
-   --moe-shared-expert-intermediate-size 5376
-   # Super-120B bottlenecks expert input/output through a 1024-dim latent.
-   # Routed experts run on moe_latent_size, NOT hidden_size, with two extra
-   # fc1/fc2 latent projections per MoE layer. The miles NemotronH bridge
-   # surfaces this from HF config; the CLI arg keeps Megatron's parser happy.
-   --moe-latent-size 1024
-   --moe-router-score-function sigmoid
-   --moe-router-enable-expert-bias
-   --moe-grouped-gemm
-   --moe-router-dtype fp32
-   # Routing: HF config has n_group=1 (MoE groups), topk_group=1,
-   # routed_scaling_factor=5.0. With n_group=1, group-limited routing is a
-   # no-op (single group of 512). `n_groups=8` in HF is Mamba groups —
-   # unrelated to MoE.
-   --moe-router-num-groups 1
-   --moe-router-group-topk 1
-   --moe-router-topk-scaling-factor 5.0
-   --moe-router-pre-softmax
-   # Match nano-30b-a3b (known-working MoE RL on nemotron_h) settings.
-   --moe-router-load-balancing-type seq_aux_loss
-   --moe-router-bias-update-rate 0
-   --moe-aux-loss-coeff 0
-)
diff --git a/scripts/models/nemotron-3-ultra-550b-a55b-4layer.py b/scripts/models/nemotron-3-ultra-550b-a55b-4layer.py
new file mode 100644
index 00000000000..c8795f3d414
--- /dev/null
+++ b/scripts/models/nemotron-3-ultra-550b-a55b-4layer.py
@@ -0,0 +1,48 @@
+# 4-layer slice of NVIDIA Nemotron-3-Ultra-550B-A55B, for single-node (8 GPU) CI.
+#
+# Built by cluster_scripts/debug_tool_set/checkpoint/prune_nemotron_h.py, which
+# keeps source layers 0,1,7,8 and renumbers them 0..3. That selection is the
+# cheapest one covering every block type the full 108-layer model has:
+#
+#   layer 0 mamba   layer 1 moe   layer 2 attention   layer 3 moe   -> "ME*E"
+#   MTP head: attention + moe                                       -> "*E"
+#
+# A prefix cut would need 8 layers to reach the first attention layer and drag
+# in 4 MoE layers (~44B params) instead of 2. Everything else (512 experts,
+# top-22, moe_latent_size=2048, sigmoid router + expert bias) is unchanged from
+# nemotron-3-ultra-550b-a55b.py, so the weight-conversion path is identical.
+
+
+def model_args() -> str:
+    return (
+        "--disable-bias-linear "
+        "--group-query-attention "
+        "--num-attention-heads 64 "
+        "--num-query-groups 2 "
+        "--kv-channels 128 "
+        "--num-layers 4 "
+        "--hidden-size 8192 "
+        "--ffn-hidden-size 5120 "
+        "--normalization RMSNorm "
+        "--position-embedding-type none "
+        "--vocab-size 131072 "
+        "--make-vocab-size-divisible-by 128 "
+        "--untie-embeddings-and-output-weights "
+        # MoE specifics (identical to the full model)
+        "--num-experts 512 "
+        "--moe-router-topk 22 "
+        "--moe-ffn-hidden-size 5120 "
+        "--moe-shared-expert-intermediate-size 10240 "
+        "--moe-latent-size 2048 "
+        "--moe-router-score-function sigmoid "
+        "--moe-router-enable-expert-bias "
+        "--moe-grouped-gemm "
+        "--moe-router-dtype fp32 "
+        "--moe-router-num-groups 1 "
+        "--moe-router-group-topk 1 "
+        "--moe-router-topk-scaling-factor 5.0 "
+        "--moe-router-pre-softmax "
+        "--moe-router-load-balancing-type seq_aux_loss "
+        "--moe-router-bias-update-rate 0 "
+        "--moe-aux-loss-coeff 0 "
+    )
diff --git a/scripts/models/nemotron-3-ultra-550b-a55b-4layer.sh b/scripts/models/nemotron-3-ultra-550b-a55b-4layer.sh
deleted file mode 100644
index fc3be08a798..00000000000
--- a/scripts/models/nemotron-3-ultra-550b-a55b-4layer.sh
+++ /dev/null
@@ -1,47 +0,0 @@
-# 4-layer slice of NVIDIA Nemotron-3-Ultra-550B-A55B, for single-node (8 GPU) CI.
-#
-# Built by cluster_scripts/debug_tool_set/checkpoint/prune_nemotron_h.py, which
-# keeps source layers 0,1,7,8 and renumbers them 0..3. That selection is the
-# cheapest one covering every block type the full 108-layer model has:
-#
-#   layer 0 mamba   layer 1 moe   layer 2 attention   layer 3 moe   -> "ME*E"
-#   MTP head: attention + moe                                       -> "*E"
-#
-# A prefix cut would need 8 layers to reach the first attention layer and drag
-# in 4 MoE layers (~44B params) instead of 2. Everything else (512 experts,
-# top-22, moe_latent_size=2048, sigmoid router + expert bias) is unchanged from
-# nemotron-3-ultra-550b-a55b.sh, so the weight-conversion path is identical.
-
-MODEL_ARGS=(
-   --disable-bias-linear
-   --group-query-attention
-   --num-attention-heads 64
-   --num-query-groups 2
-   --kv-channels 128
-   --num-layers 4
-   --hidden-size 8192
-   --ffn-hidden-size 5120
-   --normalization RMSNorm
-   --position-embedding-type none
-   --vocab-size 131072
-   --make-vocab-size-divisible-by 128
-   --untie-embeddings-and-output-weights
-
-   # MoE specifics (identical to the full model)
-   --num-experts 512
-   --moe-router-topk 22
-   --moe-ffn-hidden-size 5120
-   --moe-shared-expert-intermediate-size 10240
-   --moe-latent-size 2048
-   --moe-router-score-function sigmoid
-   --moe-router-enable-expert-bias
-   --moe-grouped-gemm
-   --moe-router-dtype fp32
-   --moe-router-num-groups 1
-   --moe-router-group-topk 1
-   --moe-router-topk-scaling-factor 5.0
-   --moe-router-pre-softmax
-   --moe-router-load-balancing-type seq_aux_loss
-   --moe-router-bias-update-rate 0
-   --moe-aux-loss-coeff 0
-)
diff --git a/scripts/models/nemotron-3-ultra-550b-a55b.py b/scripts/models/nemotron-3-ultra-550b-a55b.py
new file mode 100644
index 00000000000..d11d32eeb7d
--- /dev/null
+++ b/scripts/models/nemotron-3-ultra-550b-a55b.py
@@ -0,0 +1,52 @@
+# NVIDIA Nemotron-3-Ultra-550B-A55B (BF16, MoE nemotron_h = hybrid Mamba2 + Attention + MoE).
+# HF config:
+#   num_hidden_layers=108  hidden_size=8192  num_attention_heads=64  num_key_value_heads=2
+#   head_dim=128  intermediate_size=5120  moe_intermediate_size=5120
+#   n_routed_experts=512  num_experts_per_tok=22  n_shared_experts=1
+#   moe_shared_expert_intermediate_size=10240  routed_scaling_factor=5.0
+#   moe_latent_size=2048  n_group=1  topk_group=1  sigmoid routing + aux-free expert bias
+#   num_nextn_predict_layers=1 (MTP head)   mamba n_groups=8
+# Same AutoBridge path as Super-120B (--megatron-to-hf-mode bridge) + miles
+# NemotronHBridge MoE/latent shim (miles_plugins/megatron_bridge/nemotron_h.py).
+# NOTE: Mamba n_groups=8 forces attention/mamba tensor-parallel <= 8 (n_groups % tp == 0).
+
+
+def model_args() -> str:
+    return (
+        "--disable-bias-linear "
+        "--group-query-attention "
+        "--num-attention-heads 64 "
+        "--num-query-groups 2 "
+        "--kv-channels 128 "
+        "--num-layers 108 "
+        "--hidden-size 8192 "
+        "--ffn-hidden-size 5120 "
+        "--normalization RMSNorm "
+        "--position-embedding-type none "
+        "--vocab-size 131072 "
+        "--make-vocab-size-divisible-by 128 "
+        "--untie-embeddings-and-output-weights "
+        # MoE specifics
+        "--num-experts 512 "
+        "--moe-router-topk 22 "
+        "--moe-ffn-hidden-size 5120 "
+        "--moe-shared-expert-intermediate-size 10240 "
+        # Ultra-550B bottlenecks expert input/output through a 2048-dim latent
+        # (routed experts run on moe_latent_size, not hidden_size; two extra fc1/fc2
+        # latent projections per MoE layer). Surfaced from HF config by the miles
+        # NemotronH bridge; the CLI arg keeps Megatron's parser happy.
+        "--moe-latent-size 2048 "
+        "--moe-router-score-function sigmoid "
+        "--moe-router-enable-expert-bias "
+        "--moe-grouped-gemm "
+        "--moe-router-dtype fp32 "
+        # n_group=1 (MoE groups) -> group-limited routing is a no-op (single group of
+        # 512). HF n_groups=8 is the Mamba groups, unrelated to MoE.
+        "--moe-router-num-groups 1 "
+        "--moe-router-group-topk 1 "
+        "--moe-router-topk-scaling-factor 5.0 "
+        "--moe-router-pre-softmax "
+        "--moe-router-load-balancing-type seq_aux_loss "
+        "--moe-router-bias-update-rate 0 "
+        "--moe-aux-loss-coeff 0 "
+    )
diff --git a/scripts/models/nemotron-3-ultra-550b-a55b.sh b/scripts/models/nemotron-3-ultra-550b-a55b.sh
deleted file mode 100644
index 73d99aca60b..00000000000
--- a/scripts/models/nemotron-3-ultra-550b-a55b.sh
+++ /dev/null
@@ -1,51 +0,0 @@
-# NVIDIA Nemotron-3-Ultra-550B-A55B (BF16, MoE nemotron_h = hybrid Mamba2 + Attention + MoE).
-# HF config:
-#   num_hidden_layers=108  hidden_size=8192  num_attention_heads=64  num_key_value_heads=2
-#   head_dim=128  intermediate_size=5120  moe_intermediate_size=5120
-#   n_routed_experts=512  num_experts_per_tok=22  n_shared_experts=1
-#   moe_shared_expert_intermediate_size=10240  routed_scaling_factor=5.0
-#   moe_latent_size=2048  n_group=1  topk_group=1  sigmoid routing + aux-free expert bias
-#   num_nextn_predict_layers=1 (MTP head)   mamba n_groups=8
-# Same AutoBridge path as Super-120B (--megatron-to-hf-mode bridge) + miles
-# NemotronHBridge MoE/latent shim (miles_plugins/megatron_bridge/nemotron_h.py).
-# NOTE: Mamba n_groups=8 forces attention/mamba tensor-parallel <= 8 (n_groups % tp == 0).
-
-MODEL_ARGS=(
-   --disable-bias-linear
-   --group-query-attention
-   --num-attention-heads 64
-   --num-query-groups 2
-   --kv-channels 128
-   --num-layers 108
-   --hidden-size 8192
-   --ffn-hidden-size 5120
-   --normalization RMSNorm
-   --position-embedding-type none
-   --vocab-size 131072
-   --make-vocab-size-divisible-by 128
-   --untie-embeddings-and-output-weights
-
-   # MoE specifics
-   --num-experts 512
-   --moe-router-topk 22
-   --moe-ffn-hidden-size 5120
-   --moe-shared-expert-intermediate-size 10240
-   # Ultra-550B bottlenecks expert input/output through a 2048-dim latent
-   # (routed experts run on moe_latent_size, not hidden_size; two extra fc1/fc2
-   # latent projections per MoE layer). Surfaced from HF config by the miles
-   # NemotronH bridge; the CLI arg keeps Megatron's parser happy.
-   --moe-latent-size 2048
-   --moe-router-score-function sigmoid
-   --moe-router-enable-expert-bias
-   --moe-grouped-gemm
-   --moe-router-dtype fp32
-   # n_group=1 (MoE groups) -> group-limited routing is a no-op (single group of
-   # 512). HF n_groups=8 is the Mamba groups, unrelated to MoE.
-   --moe-router-num-groups 1
-   --moe-router-group-topk 1
-   --moe-router-topk-scaling-factor 5.0
-   --moe-router-pre-softmax
-   --moe-router-load-balancing-type seq_aux_loss
-   --moe-router-bias-update-rate 0
-   --moe-aux-loss-coeff 0
-)
diff --git a/scripts/models/qwen2.5-0.5B.py b/scripts/models/qwen2.5-0.5B.py
new file mode 100644
index 00000000000..d1581103135
--- /dev/null
+++ b/scripts/models/qwen2.5-0.5B.py
@@ -0,0 +1,17 @@
+def model_args() -> str:
+    return (
+        "--swiglu "
+        "--num-layers 24 "
+        "--hidden-size 896 "
+        "--ffn-hidden-size 4864 "
+        "--num-attention-heads 14 "
+        "--use-rotary-position-embeddings "
+        "--disable-bias-linear "
+        "--add-qkv-bias "
+        "--normalization RMSNorm "
+        "--norm-epsilon 1e-6 "
+        "--rotary-base 1000000 "
+        "--group-query-attention "
+        "--num-query-groups 2 "
+        "--vocab-size 151936 "
+    )
diff --git a/scripts/models/qwen2.5-0.5B.sh b/scripts/models/qwen2.5-0.5B.sh
deleted file mode 100644
index 66d5b29a024..00000000000
--- a/scripts/models/qwen2.5-0.5B.sh
+++ /dev/null
@@ -1,16 +0,0 @@
-MODEL_ARGS=(
-   --swiglu
-   --num-layers 24
-   --hidden-size 896
-   --ffn-hidden-size 4864
-   --num-attention-heads 14
-   --use-rotary-position-embeddings
-   --disable-bias-linear
-   --add-qkv-bias
-   --normalization "RMSNorm"
-   --norm-epsilon 1e-6
-   --rotary-base 1000000
-   --group-query-attention
-   --num-query-groups 2
-   --vocab-size 151936
-)
\ No newline at end of file
diff --git a/scripts/models/qwen2.5-1.5B.py b/scripts/models/qwen2.5-1.5B.py
new file mode 100644
index 00000000000..19d95e69dbc
--- /dev/null
+++ b/scripts/models/qwen2.5-1.5B.py
@@ -0,0 +1,17 @@
+def model_args() -> str:
+    return (
+        "--swiglu "
+        "--num-layers 28 "
+        "--hidden-size 1536 "
+        "--ffn-hidden-size 8960 "
+        "--num-attention-heads 12 "
+        "--use-rotary-position-embeddings "
+        "--disable-bias-linear "
+        "--add-qkv-bias "
+        "--normalization RMSNorm "
+        "--norm-epsilon 1e-6 "
+        "--rotary-base 10000 "
+        "--group-query-attention "
+        "--num-query-groups 2 "
+        "--vocab-size 151936 "
+    )
diff --git a/scripts/models/qwen2.5-1.5B.sh b/scripts/models/qwen2.5-1.5B.sh
deleted file mode 100644
index b046a95c66b..00000000000
--- a/scripts/models/qwen2.5-1.5B.sh
+++ /dev/null
@@ -1,16 +0,0 @@
-MODEL_ARGS=(
-   --swiglu
-   --num-layers 28
-   --hidden-size 1536
-   --ffn-hidden-size 8960
-   --num-attention-heads 12
-   --use-rotary-position-embeddings
-   --disable-bias-linear
-   --add-qkv-bias
-   --normalization "RMSNorm"
-   --norm-epsilon 1e-6
-   --rotary-base 10000
-   --group-query-attention
-   --num-query-groups 2
-   --vocab-size 151936
-)
\ No newline at end of file
diff --git a/scripts/models/qwen2.5-32B.py b/scripts/models/qwen2.5-32B.py
new file mode 100644
index 00000000000..bc1251d5ccf
--- /dev/null
+++ b/scripts/models/qwen2.5-32B.py
@@ -0,0 +1,18 @@
+def model_args() -> str:
+    return (
+        "--swiglu "
+        "--num-layers 64 "
+        "--hidden-size 5120 "
+        "--ffn-hidden-size 27648 "
+        "--num-attention-heads 40 "
+        "--group-query-attention "
+        "--num-query-groups 8 "
+        "--use-rotary-position-embeddings "
+        "--disable-bias-linear "
+        "--add-qkv-bias "
+        "--normalization RMSNorm "
+        "--norm-epsilon 1e-5 "
+        "--rotary-base 1000000 "
+        "--vocab-size 152064 "
+        "--untie-embeddings-and-output-weights "
+    )
diff --git a/scripts/models/qwen2.5-32B.sh b/scripts/models/qwen2.5-32B.sh
deleted file mode 100644
index 26b49845a41..00000000000
--- a/scripts/models/qwen2.5-32B.sh
+++ /dev/null
@@ -1,17 +0,0 @@
-MODEL_ARGS=(
-   --swiglu
-   --num-layers 64
-   --hidden-size 5120
-   --ffn-hidden-size 27648
-   --num-attention-heads 40
-   --group-query-attention
-   --num-query-groups 8
-   --use-rotary-position-embeddings
-   --disable-bias-linear
-   --add-qkv-bias
-   --normalization "RMSNorm"
-   --norm-epsilon 1e-5
-   --rotary-base 1000000
-   --vocab-size 152064
-   --untie-embeddings-and-output-weights
-)
\ No newline at end of file
diff --git a/scripts/models/qwen2.5-3B.py b/scripts/models/qwen2.5-3B.py
new file mode 100644
index 00000000000..fe922621ed0
--- /dev/null
+++ b/scripts/models/qwen2.5-3B.py
@@ -0,0 +1,17 @@
+def model_args() -> str:
+    return (
+        "--swiglu "
+        "--num-layers 36 "
+        "--hidden-size 2048 "
+        "--ffn-hidden-size 11008 "
+        "--num-attention-heads 16 "
+        "--use-rotary-position-embeddings "
+        "--disable-bias-linear "
+        "--add-qkv-bias "
+        "--normalization RMSNorm "
+        "--norm-epsilon 1e-6 "
+        "--rotary-base 1000000 "
+        "--group-query-attention "
+        "--num-query-groups 2 "
+        "--vocab-size 151936 "
+    )
diff --git a/scripts/models/qwen2.5-3B.sh b/scripts/models/qwen2.5-3B.sh
deleted file mode 100644
index 9da5a9e0339..00000000000
--- a/scripts/models/qwen2.5-3B.sh
+++ /dev/null
@@ -1,16 +0,0 @@
-MODEL_ARGS=(
-   --swiglu
-   --num-layers 36
-   --hidden-size 2048
-   --ffn-hidden-size 11008
-   --num-attention-heads 16
-   --use-rotary-position-embeddings
-   --disable-bias-linear
-   --add-qkv-bias
-   --normalization "RMSNorm"
-   --norm-epsilon 1e-6
-   --rotary-base 1000000
-   --group-query-attention
-   --num-query-groups 2
-   --vocab-size 151936
-)
diff --git a/scripts/models/qwen2.5-7B.py b/scripts/models/qwen2.5-7B.py
new file mode 100644
index 00000000000..7fcbb618cff
--- /dev/null
+++ b/scripts/models/qwen2.5-7B.py
@@ -0,0 +1,18 @@
+def model_args() -> str:
+    return (
+        "--swiglu "
+        "--num-layers 28 "
+        "--hidden-size 3584 "
+        "--ffn-hidden-size 18944 "
+        "--num-attention-heads 28 "
+        "--group-query-attention "
+        "--num-query-groups 4 "
+        "--use-rotary-position-embeddings "
+        "--disable-bias-linear "
+        "--add-qkv-bias "
+        "--normalization RMSNorm "
+        "--norm-epsilon 1e-06 "
+        "--rotary-base 1000000 "
+        "--vocab-size 152064 "
+        "--untie-embeddings-and-output-weights "
+    )
diff --git a/scripts/models/qwen2.5-7B.sh b/scripts/models/qwen2.5-7B.sh
deleted file mode 100644
index eba912b1d7c..00000000000
--- a/scripts/models/qwen2.5-7B.sh
+++ /dev/null
@@ -1,17 +0,0 @@
-MODEL_ARGS=(
-   --swiglu
-   --num-layers 28
-   --hidden-size 3584
-   --ffn-hidden-size 18944
-   --num-attention-heads 28
-   --group-query-attention
-   --num-query-groups 4
-   --use-rotary-position-embeddings
-   --disable-bias-linear
-   --add-qkv-bias
-   --normalization "RMSNorm"
-   --norm-epsilon 1e-06
-   --rotary-base 1000000
-   --vocab-size 152064
-   --untie-embeddings-and-output-weights
-)
\ No newline at end of file
diff --git a/scripts/models/qwen3-0.6B.py b/scripts/models/qwen3-0.6B.py
new file mode 100644
index 00000000000..781fe2bd437
--- /dev/null
+++ b/scripts/models/qwen3-0.6B.py
@@ -0,0 +1,18 @@
+def model_args() -> str:
+    return (
+        "--swiglu "
+        "--num-layers 28 "
+        "--hidden-size 1024 "
+        "--ffn-hidden-size 3072 "
+        "--num-attention-heads 16 "
+        "--group-query-attention "
+        "--num-query-groups 8 "
+        "--use-rotary-position-embeddings "
+        "--disable-bias-linear "
+        "--normalization RMSNorm "
+        "--norm-epsilon 1e-6 "
+        "--rotary-base 1000000 "
+        "--vocab-size 151936 "
+        "--kv-channels 128 "
+        "--qk-layernorm "
+    )
diff --git a/scripts/models/qwen3-0.6B.sh b/scripts/models/qwen3-0.6B.sh
deleted file mode 100644
index f484ec9519b..00000000000
--- a/scripts/models/qwen3-0.6B.sh
+++ /dev/null
@@ -1,17 +0,0 @@
-MODEL_ARGS=(
-   --swiglu
-   --num-layers 28
-   --hidden-size 1024
-   --ffn-hidden-size 3072
-   --num-attention-heads 16
-   --group-query-attention
-   --num-query-groups 8
-   --use-rotary-position-embeddings
-   --disable-bias-linear
-   --normalization "RMSNorm"
-   --norm-epsilon 1e-6
-   --rotary-base 1000000
-   --vocab-size 151936
-   --kv-channels 128
-   --qk-layernorm
-)
\ No newline at end of file
diff --git a/scripts/models/qwen3-1.7B.py b/scripts/models/qwen3-1.7B.py
new file mode 100644
index 00000000000..e74dfac0f56
--- /dev/null
+++ b/scripts/models/qwen3-1.7B.py
@@ -0,0 +1,22 @@
+import os
+
+
+def model_args(rotary_base: str | None = None) -> str:
+    rotary_base = rotary_base if rotary_base is not None else os.environ.get("MODEL_ARGS_ROTARY_BASE") or "1000000"
+    return (
+        "--swiglu "
+        "--num-layers 28 "
+        "--hidden-size 2048 "
+        "--ffn-hidden-size 6144 "
+        "--num-attention-heads 16 "
+        "--group-query-attention "
+        "--num-query-groups 8 "
+        "--use-rotary-position-embeddings "
+        "--disable-bias-linear "
+        "--normalization RMSNorm "
+        "--norm-epsilon 1e-6 "
+        f"--rotary-base {rotary_base} "
+        "--vocab-size 151936 "
+        "--kv-channels 128 "
+        "--qk-layernorm "
+    )
diff --git a/scripts/models/qwen3-1.7B.sh b/scripts/models/qwen3-1.7B.sh
deleted file mode 100644
index 7435996337e..00000000000
--- a/scripts/models/qwen3-1.7B.sh
+++ /dev/null
@@ -1,17 +0,0 @@
-MODEL_ARGS=(
-   --swiglu
-   --num-layers 28
-   --hidden-size 2048
-   --ffn-hidden-size 6144
-   --num-attention-heads 16
-   --group-query-attention
-   --num-query-groups 8
-   --use-rotary-position-embeddings
-   --disable-bias-linear
-   --normalization "RMSNorm"
-   --norm-epsilon 1e-6
-   --rotary-base "${MODEL_ARGS_ROTARY_BASE:-1000000}"
-   --vocab-size 151936
-   --kv-channels 128
-   --qk-layernorm
-)
\ No newline at end of file
diff --git a/scripts/models/qwen3-14B.py b/scripts/models/qwen3-14B.py
new file mode 100644
index 00000000000..12b0af4bd78
--- /dev/null
+++ b/scripts/models/qwen3-14B.py
@@ -0,0 +1,19 @@
+def model_args() -> str:
+    return (
+        "--swiglu "
+        "--num-layers 40 "
+        "--hidden-size 5120 "
+        "--ffn-hidden-size 17408 "
+        "--num-attention-heads 40 "
+        "--group-query-attention "
+        "--num-query-groups 8 "
+        "--use-rotary-position-embeddings "
+        "--disable-bias-linear "
+        "--normalization RMSNorm "
+        "--norm-epsilon 1e-6 "
+        "--rotary-base 1000000 "
+        "--vocab-size 151936 "
+        "--kv-channels 128 "
+        "--qk-layernorm "
+        "--untie-embeddings-and-output-weights "
+    )
diff --git a/scripts/models/qwen3-14B.sh b/scripts/models/qwen3-14B.sh
deleted file mode 100644
index 11b9377da00..00000000000
--- a/scripts/models/qwen3-14B.sh
+++ /dev/null
@@ -1,18 +0,0 @@
-MODEL_ARGS=(
-   --swiglu
-   --num-layers 40
-   --hidden-size 5120
-   --ffn-hidden-size 17408
-   --num-attention-heads 40
-   --group-query-attention
-   --num-query-groups 8
-   --use-rotary-position-embeddings
-   --disable-bias-linear
-   --normalization "RMSNorm"
-   --norm-epsilon 1e-6
-   --rotary-base 1000000
-   --vocab-size 151936
-   --kv-channels 128
-   --qk-layernorm
-   --untie-embeddings-and-output-weights
-)
diff --git a/scripts/models/qwen3-235B-A22B.py b/scripts/models/qwen3-235B-A22B.py
new file mode 100644
index 00000000000..70c53786ae0
--- /dev/null
+++ b/scripts/models/qwen3-235B-A22B.py
@@ -0,0 +1,42 @@
+import os
+
+from model_args_utils import moe_layer_freq
+
+
+NLAYERS = 94
+FIRST_K_DENSE_REPLACE = 0
+
+
+def model_args(rotary_base: str | None = None) -> str:
+    rotary_base = rotary_base if rotary_base is not None else os.environ.get("MODEL_ARGS_ROTARY_BASE") or "1000000"
+    return (
+        "--disable-bias-linear "
+        "--qk-layernorm "
+        "--group-query-attention "
+        "--num-attention-heads 64 "
+        "--num-query-groups 4 "
+        "--kv-channels 128 "
+        "--num-layers 94 "
+        "--hidden-size 4096 "
+        "--ffn-hidden-size 12288 "
+        "--normalization RMSNorm "
+        "--position-embedding-type rope "
+        "--norm-epsilon 1e-6 "
+        "--rotary-percent 1.0 "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--vocab-size 151936 "
+        f"--rotary-base {rotary_base} "
+        # moe
+        "--moe-ffn-hidden-size 1536 "
+        "--moe-router-score-function softmax "
+        "--moe-token-dispatcher-type alltoall "
+        "--moe-router-topk 8 "
+        f"--moe-layer-freq {moe_layer_freq(nlayers=NLAYERS, first_k_dense_replace=FIRST_K_DENSE_REPLACE)} "
+        "--num-experts 128 "
+        "--moe-grouped-gemm "
+        "--moe-token-drop-policy probs "
+        "--moe-router-dtype fp32 "
+        "--moe-permute-fusion "
+        "--moe-aux-loss-coeff 0 "
+    )
diff --git a/scripts/models/qwen3-235B-A22B.sh b/scripts/models/qwen3-235B-A22B.sh
deleted file mode 100644
index 1f663552653..00000000000
--- a/scripts/models/qwen3-235B-A22B.sh
+++ /dev/null
@@ -1,49 +0,0 @@
-# qwen3-235B-a22B
-NLAYERS=94
-FIRST_K_DENSE_REPLACE=0
-
-arr=()
-for ((i=0; i<NLAYERS; i++)); do
-  if (( i < FIRST_K_DENSE_REPLACE )); then
-    arr+=(0)
-  else
-    arr+=(1)
-  fi
-done
-
-printf -v MOE_LAYER_FREQ "[%s]" "$(IFS=', '; echo "${arr[*]}")"
-
-MODEL_ARGS=(
-   --disable-bias-linear
-   --qk-layernorm
-   --group-query-attention
-   --num-attention-heads 64
-   --num-query-groups 4
-   --kv-channels 128
-   --num-layers 94
-   --hidden-size 4096
-   --ffn-hidden-size 12288
-
-   --normalization RMSNorm
-   --position-embedding-type rope
-   --norm-epsilon 1e-6
-   --rotary-percent 1.0
-   --swiglu
-   --untie-embeddings-and-output-weights
-   --vocab-size 151936
-
-   --rotary-base "${MODEL_ARGS_ROTARY_BASE:-1000000}"
-
-   # moe
-   --moe-ffn-hidden-size 1536
-   --moe-router-score-function softmax
-   --moe-token-dispatcher-type alltoall
-   --moe-router-topk 8
-   --moe-layer-freq "$MOE_LAYER_FREQ"
-   --num-experts 128
-   --moe-grouped-gemm
-   --moe-token-drop-policy probs
-   --moe-router-dtype fp32
-   --moe-permute-fusion
-   --moe-aux-loss-coeff 0
-)
\ No newline at end of file
diff --git a/scripts/models/qwen3-30B-A3B-5layer.py b/scripts/models/qwen3-30B-A3B-5layer.py
new file mode 100644
index 00000000000..6686fb29056
--- /dev/null
+++ b/scripts/models/qwen3-30B-A3B-5layer.py
@@ -0,0 +1,5 @@
+from model_args_utils import load_sibling_model_args
+
+
+def model_args() -> str:
+    return load_sibling_model_args(__file__, "qwen3-30B-A3B", nlayers=5)
diff --git a/scripts/models/qwen3-30B-A3B-5layer.sh b/scripts/models/qwen3-30B-A3B-5layer.sh
deleted file mode 100644
index 449461ae208..00000000000
--- a/scripts/models/qwen3-30B-A3B-5layer.sh
+++ /dev/null
@@ -1 +0,0 @@
-MODEL_ARGS_NUM_LAYERS=5 source "$(dirname -- "${BASH_SOURCE[0]}")/qwen3-30B-A3B.sh"
diff --git a/scripts/models/qwen3-30B-A3B.py b/scripts/models/qwen3-30B-A3B.py
new file mode 100644
index 00000000000..4f1a4c33010
--- /dev/null
+++ b/scripts/models/qwen3-30B-A3B.py
@@ -0,0 +1,42 @@
+import os
+
+from model_args_utils import moe_layer_freq
+
+
+FIRST_K_DENSE_REPLACE = 0
+
+
+def model_args(nlayers: int | None = None, rotary_base: str | None = None) -> str:
+    nlayers = nlayers if nlayers is not None else int(os.environ.get("MODEL_ARGS_NUM_LAYERS") or 48)
+    rotary_base = rotary_base if rotary_base is not None else os.environ.get("MODEL_ARGS_ROTARY_BASE") or "1000000"
+    return (
+        "--disable-bias-linear "
+        "--qk-layernorm "
+        "--group-query-attention "
+        "--num-attention-heads 32 "
+        "--num-query-groups 4 "
+        "--kv-channels 128 "
+        f"--num-layers {nlayers} "
+        "--hidden-size 2048 "
+        "--ffn-hidden-size 6144 "
+        "--normalization RMSNorm "
+        "--position-embedding-type rope "
+        "--norm-epsilon 1e-6 "
+        "--rotary-percent 1.0 "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--vocab-size 151936 "
+        f"--rotary-base {rotary_base} "
+        # moe
+        "--moe-ffn-hidden-size 768 "
+        "--moe-router-score-function softmax "
+        "--moe-token-dispatcher-type alltoall "
+        "--moe-router-topk 8 "
+        f"--moe-layer-freq {moe_layer_freq(nlayers=nlayers, first_k_dense_replace=FIRST_K_DENSE_REPLACE)} "
+        "--num-experts 128 "
+        "--moe-grouped-gemm "
+        "--moe-token-drop-policy probs "
+        "--moe-router-dtype fp32 "
+        "--moe-permute-fusion "
+        "--moe-aux-loss-coeff 0 "
+    )
diff --git a/scripts/models/qwen3-30B-A3B.sh b/scripts/models/qwen3-30B-A3B.sh
deleted file mode 100644
index 0221af9bba7..00000000000
--- a/scripts/models/qwen3-30B-A3B.sh
+++ /dev/null
@@ -1,49 +0,0 @@
-NLAYERS="${MODEL_ARGS_NUM_LAYERS:-48}"
-FIRST_K_DENSE_REPLACE=0
-
-arr=()
-for ((i=0; i<NLAYERS; i++)); do
-  if (( i < FIRST_K_DENSE_REPLACE )); then
-    arr+=(0)
-  else
-    arr+=(1)
-  fi
-done
-
-printf -v MOE_LAYER_FREQ "[%s]" "$(IFS=', '; echo "${arr[*]}")"
-
-
-MODEL_ARGS=(
-   --disable-bias-linear
-   --qk-layernorm
-   --group-query-attention
-   --num-attention-heads 32
-   --num-query-groups 4
-   --kv-channels 128
-   --num-layers $NLAYERS
-   --hidden-size 2048
-   --ffn-hidden-size 6144
-
-   --normalization RMSNorm
-   --position-embedding-type rope
-   --norm-epsilon 1e-6
-   --rotary-percent 1.0
-   --swiglu
-   --untie-embeddings-and-output-weights
-   --vocab-size 151936
-
-   --rotary-base "${MODEL_ARGS_ROTARY_BASE:-1000000}"
-
-   # moe
-   --moe-ffn-hidden-size 768
-   --moe-router-score-function softmax
-   --moe-token-dispatcher-type alltoall
-   --moe-router-topk 8
-   --moe-layer-freq "$MOE_LAYER_FREQ"
-   --num-experts 128
-   --moe-grouped-gemm
-   --moe-token-drop-policy probs
-   --moe-router-dtype fp32
-   --moe-permute-fusion
-   --moe-aux-loss-coeff 0
-)
\ No newline at end of file
diff --git a/scripts/models/qwen3-32B.py b/scripts/models/qwen3-32B.py
new file mode 100644
index 00000000000..218e80450c9
--- /dev/null
+++ b/scripts/models/qwen3-32B.py
@@ -0,0 +1,19 @@
+def model_args() -> str:
+    return (
+        "--swiglu "
+        "--num-layers 64 "
+        "--hidden-size 5120 "
+        "--ffn-hidden-size 25600 "
+        "--num-attention-heads 64 "
+        "--group-query-attention "
+        "--num-query-groups 8 "
+        "--use-rotary-position-embeddings "
+        "--disable-bias-linear "
+        "--normalization RMSNorm "
+        "--norm-epsilon 1e-6 "
+        "--rotary-base 1000000 "
+        "--vocab-size 151936 "
+        "--kv-channels 128 "
+        "--qk-layernorm "
+        "--untie-embeddings-and-output-weights "
+    )
diff --git a/scripts/models/qwen3-32B.sh b/scripts/models/qwen3-32B.sh
deleted file mode 100644
index e7407e327c9..00000000000
--- a/scripts/models/qwen3-32B.sh
+++ /dev/null
@@ -1,18 +0,0 @@
-MODEL_ARGS=(
-   --swiglu
-   --num-layers 64
-   --hidden-size 5120
-   --ffn-hidden-size 25600
-   --num-attention-heads 64
-   --group-query-attention
-   --num-query-groups 8
-   --use-rotary-position-embeddings
-   --disable-bias-linear
-   --normalization "RMSNorm"
-   --norm-epsilon 1e-6
-   --rotary-base 1000000
-   --vocab-size 151936
-   --kv-channels 128
-   --qk-layernorm
-   --untie-embeddings-and-output-weights
-)
diff --git a/scripts/models/qwen3-4B-Instruct-2507.py b/scripts/models/qwen3-4B-Instruct-2507.py
new file mode 100644
index 00000000000..7c8150db382
--- /dev/null
+++ b/scripts/models/qwen3-4B-Instruct-2507.py
@@ -0,0 +1,5 @@
+from model_args_utils import load_sibling_model_args
+
+
+def model_args() -> str:
+    return load_sibling_model_args(__file__, "qwen3-4B", rotary_base=5000000)
diff --git a/scripts/models/qwen3-4B-Instruct-2507.sh b/scripts/models/qwen3-4B-Instruct-2507.sh
deleted file mode 100644
index 67d13c0c823..00000000000
--- a/scripts/models/qwen3-4B-Instruct-2507.sh
+++ /dev/null
@@ -1 +0,0 @@
-MODEL_ARGS_ROTARY_BASE=5000000 source "$(dirname -- "${BASH_SOURCE[0]}")/qwen3-4B.sh"
\ No newline at end of file
diff --git a/scripts/models/qwen3-4B.py b/scripts/models/qwen3-4B.py
new file mode 100644
index 00000000000..b1a40729b8b
--- /dev/null
+++ b/scripts/models/qwen3-4B.py
@@ -0,0 +1,22 @@
+import os
+
+
+def model_args(rotary_base: str | None = None) -> str:
+    rotary_base = rotary_base if rotary_base is not None else os.environ.get("MODEL_ARGS_ROTARY_BASE") or "1000000"
+    return (
+        "--swiglu "
+        "--num-layers 36 "
+        "--hidden-size 2560 "
+        "--ffn-hidden-size 9728 "
+        "--num-attention-heads 32 "
+        "--group-query-attention "
+        "--num-query-groups 8 "
+        "--use-rotary-position-embeddings "
+        "--disable-bias-linear "
+        "--normalization RMSNorm "
+        "--norm-epsilon 1e-6 "
+        f"--rotary-base {rotary_base} "
+        "--vocab-size 151936 "
+        "--kv-channels 128 "
+        "--qk-layernorm "
+    )
diff --git a/scripts/models/qwen3-4B.sh b/scripts/models/qwen3-4B.sh
deleted file mode 100644
index 51f9e47581e..00000000000
--- a/scripts/models/qwen3-4B.sh
+++ /dev/null
@@ -1,17 +0,0 @@
-MODEL_ARGS=(
-   --swiglu
-   --num-layers 36
-   --hidden-size 2560
-   --ffn-hidden-size 9728
-   --num-attention-heads 32
-   --group-query-attention
-   --num-query-groups 8
-   --use-rotary-position-embeddings
-   --disable-bias-linear
-   --normalization "RMSNorm"
-   --norm-epsilon 1e-6
-   --rotary-base "${MODEL_ARGS_ROTARY_BASE:-1000000}"
-   --vocab-size 151936
-   --kv-channels 128
-   --qk-layernorm
-)
\ No newline at end of file
diff --git a/scripts/models/qwen3-8B.py b/scripts/models/qwen3-8B.py
new file mode 100644
index 00000000000..029133f0d8b
--- /dev/null
+++ b/scripts/models/qwen3-8B.py
@@ -0,0 +1,23 @@
+import os
+
+
+def model_args(rotary_base: str | None = None) -> str:
+    rotary_base = rotary_base if rotary_base is not None else os.environ.get("MODEL_ARGS_ROTARY_BASE") or "1000000"
+    return (
+        "--swiglu "
+        "--num-layers 36 "
+        "--hidden-size 4096 "
+        "--ffn-hidden-size 12288 "
+        "--num-attention-heads 32 "
+        "--group-query-attention "
+        "--num-query-groups 8 "
+        "--use-rotary-position-embeddings "
+        "--disable-bias-linear "
+        "--normalization RMSNorm "
+        "--norm-epsilon 1e-6 "
+        f"--rotary-base {rotary_base} "
+        "--vocab-size 151936 "
+        "--kv-channels 128 "
+        "--qk-layernorm "
+        "--untie-embeddings-and-output-weights "
+    )
diff --git a/scripts/models/qwen3-8B.sh b/scripts/models/qwen3-8B.sh
deleted file mode 100644
index fc573adb37b..00000000000
--- a/scripts/models/qwen3-8B.sh
+++ /dev/null
@@ -1,18 +0,0 @@
-MODEL_ARGS=(
-   --swiglu
-   --num-layers 36
-   --hidden-size 4096
-   --ffn-hidden-size 12288
-   --num-attention-heads 32
-   --group-query-attention
-   --num-query-groups 8
-   --use-rotary-position-embeddings
-   --disable-bias-linear
-   --normalization "RMSNorm"
-   --norm-epsilon 1e-6
-   --rotary-base "${MODEL_ARGS_ROTARY_BASE:-1000000}"
-   --vocab-size 151936
-   --kv-channels 128
-   --qk-layernorm
-   --untie-embeddings-and-output-weights
-)
\ No newline at end of file
diff --git a/scripts/models/qwen3-next-80B-A3B.py b/scripts/models/qwen3-next-80B-A3B.py
new file mode 100644
index 00000000000..402b870c738
--- /dev/null
+++ b/scripts/models/qwen3-next-80B-A3B.py
@@ -0,0 +1,46 @@
+from model_args_utils import moe_layer_freq
+
+
+NLAYERS = 48
+FIRST_K_DENSE_REPLACE = 0
+
+
+def model_args() -> str:
+    return (
+        "--spec miles_plugins.models.qwen3_next get_qwen3_next_spec "
+        "--disable-bias-linear "
+        "--qk-layernorm "
+        "--group-query-attention "
+        "--num-attention-heads 16 "
+        "--num-query-groups 2 "
+        "--kv-channels 256 "
+        "--num-layers 48 "
+        "--hidden-size 2048 "
+        "--ffn-hidden-size 5120 "
+        "--normalization RMSNorm "
+        "--apply-layernorm-1p "
+        "--position-embedding-type rope "
+        "--norm-epsilon 1e-6 "
+        "--rotary-percent 0.25 "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--vocab-size 151936 "
+        "--rotary-base 10000000 "
+        # moe
+        "--moe-ffn-hidden-size 512 "
+        "--moe-shared-expert-intermediate-size 512 "
+        "--moe-router-score-function softmax "
+        "--moe-token-dispatcher-type alltoall "
+        "--moe-router-topk 10 "
+        f"--moe-layer-freq {moe_layer_freq(nlayers=NLAYERS, first_k_dense_replace=FIRST_K_DENSE_REPLACE)} "
+        "--num-experts 512 "
+        "--moe-grouped-gemm "
+        "--moe-token-drop-policy probs "
+        "--moe-router-dtype fp32 "
+        "--moe-permute-fusion "
+        "--moe-aux-loss-coeff 0 "
+        # qwen3 specific
+        "--attention-output-gate "
+        "--moe-shared-expert-gate "
+        "--mtp-num-layers 1 "
+    )
diff --git a/scripts/models/qwen3-next-80B-A3B.sh b/scripts/models/qwen3-next-80B-A3B.sh
deleted file mode 100644
index e5390854126..00000000000
--- a/scripts/models/qwen3-next-80B-A3B.sh
+++ /dev/null
@@ -1,58 +0,0 @@
-NLAYERS=48
-FIRST_K_DENSE_REPLACE=0
-
-arr=()
-for ((i=0; i<NLAYERS; i++)); do
-  if (( i < FIRST_K_DENSE_REPLACE )); then
-    arr+=(0)
-  else
-    arr+=(1)
-  fi
-done
-
-printf -v MOE_LAYER_FREQ "[%s]" "$(IFS=', '; echo "${arr[*]}")"
-
-
-MODEL_ARGS=(
-   --spec "miles_plugins.models.qwen3_next" "get_qwen3_next_spec"
-
-   --disable-bias-linear
-   --qk-layernorm
-   --group-query-attention
-   --num-attention-heads 16
-   --num-query-groups 2
-   --kv-channels 256
-   --num-layers 48
-   --hidden-size 2048
-   --ffn-hidden-size 5120
-
-   --normalization RMSNorm
-   --apply-layernorm-1p
-   --position-embedding-type rope
-   --norm-epsilon 1e-6
-   --rotary-percent 0.25
-   --swiglu
-   --untie-embeddings-and-output-weights
-   --vocab-size 151936
-
-   --rotary-base 10000000
-
-   # moe
-   --moe-ffn-hidden-size 512
-   --moe-shared-expert-intermediate-size 512
-   --moe-router-score-function softmax
-   --moe-token-dispatcher-type alltoall
-   --moe-router-topk 10
-   --moe-layer-freq "$MOE_LAYER_FREQ"
-   --num-experts 512
-   --moe-grouped-gemm
-   --moe-token-drop-policy probs
-   --moe-router-dtype fp32
-   --moe-permute-fusion
-   --moe-aux-loss-coeff 0
-
-   # qwen3 specific
-   --attention-output-gate
-   --moe-shared-expert-gate
-   --mtp-num-layers 1
-)
\ No newline at end of file
diff --git a/scripts/models/qwen3.5-27B.py b/scripts/models/qwen3.5-27B.py
new file mode 100644
index 00000000000..ad57e163b99
--- /dev/null
+++ b/scripts/models/qwen3.5-27B.py
@@ -0,0 +1,24 @@
+def model_args() -> str:
+    return (
+        "--spec miles_plugins.models.qwen3_5 get_qwen3_5_spec "
+        "--disable-bias-linear "
+        "--qk-layernorm "
+        "--group-query-attention "
+        "--num-attention-heads 24 "
+        "--num-query-groups 4 "
+        "--kv-channels 256 "
+        "--num-layers 64 "
+        "--hidden-size 5120 "
+        "--ffn-hidden-size 17408 "
+        "--normalization RMSNorm "
+        "--apply-layernorm-1p "
+        "--position-embedding-type rope "
+        "--norm-epsilon 1e-6 "
+        "--rotary-percent 0.25 "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--vocab-size 248320 "
+        "--rotary-base 10000000 "
+        # qwen3.5 specific
+        "--attention-output-gate "
+    )
diff --git a/scripts/models/qwen3.5-27B.sh b/scripts/models/qwen3.5-27B.sh
deleted file mode 100644
index 5e76a6d3f9e..00000000000
--- a/scripts/models/qwen3.5-27B.sh
+++ /dev/null
@@ -1,27 +0,0 @@
-MODEL_ARGS=(
-   --spec "miles_plugins.models.qwen3_5" "get_qwen3_5_spec"
-
-   --disable-bias-linear
-   --qk-layernorm
-   --group-query-attention
-   --num-attention-heads 24
-   --num-query-groups 4
-   --kv-channels 256
-   --num-layers 64
-   --hidden-size 5120
-   --ffn-hidden-size 17408
-
-   --normalization RMSNorm
-   --apply-layernorm-1p
-   --position-embedding-type rope
-   --norm-epsilon 1e-6
-   --rotary-percent 0.25
-   --swiglu
-   --untie-embeddings-and-output-weights
-   --vocab-size 248320
-
-   --rotary-base 10000000
-
-   # qwen3.5 specific
-   --attention-output-gate
-)
diff --git a/scripts/models/qwen3.5-35B-A3B.py b/scripts/models/qwen3.5-35B-A3B.py
new file mode 100644
index 00000000000..e63ef67ee33
--- /dev/null
+++ b/scripts/models/qwen3.5-35B-A3B.py
@@ -0,0 +1,46 @@
+from model_args_utils import moe_layer_freq
+
+
+NLAYERS = 40
+FIRST_K_DENSE_REPLACE = 0
+
+
+def model_args() -> str:
+    return (
+        "--spec miles_plugins.models.qwen3_5 get_qwen3_5_spec "
+        "--disable-bias-linear "
+        "--qk-layernorm "
+        "--group-query-attention "
+        "--num-attention-heads 16 "
+        "--num-query-groups 2 "
+        "--kv-channels 256 "
+        "--num-layers 40 "
+        "--hidden-size 2048 "
+        "--ffn-hidden-size 512 "
+        "--normalization RMSNorm "
+        "--apply-layernorm-1p "
+        "--position-embedding-type rope "
+        "--norm-epsilon 1e-6 "
+        "--rotary-percent 0.25 "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--vocab-size 248320 "
+        "--rotary-base 10000000 "
+        # moe
+        "--moe-ffn-hidden-size 512 "
+        "--moe-shared-expert-intermediate-size 512 "
+        "--moe-router-score-function softmax "
+        "--moe-token-dispatcher-type alltoall "
+        "--moe-router-topk 8 "
+        f"--moe-layer-freq {moe_layer_freq(nlayers=NLAYERS, first_k_dense_replace=FIRST_K_DENSE_REPLACE)} "
+        "--num-experts 256 "
+        "--moe-grouped-gemm "
+        "--moe-token-drop-policy probs "
+        "--moe-router-dtype fp32 "
+        "--moe-permute-fusion "
+        "--moe-aux-loss-coeff 0 "
+        # qwen3.5 specific
+        "--attention-output-gate "
+        "--moe-shared-expert-gate "
+        "--mtp-num-layers 1 "
+    )
diff --git a/scripts/models/qwen3.5-35B-A3B.sh b/scripts/models/qwen3.5-35B-A3B.sh
deleted file mode 100644
index e6912b17dd8..00000000000
--- a/scripts/models/qwen3.5-35B-A3B.sh
+++ /dev/null
@@ -1,58 +0,0 @@
-NLAYERS=40
-FIRST_K_DENSE_REPLACE=0
-
-arr=()
-for ((i=0; i<NLAYERS; i++)); do
-  if (( i < FIRST_K_DENSE_REPLACE )); then
-    arr+=(0)
-  else
-    arr+=(1)
-  fi
-done
-
-printf -v MOE_LAYER_FREQ "[%s]" "$(IFS=', '; echo "${arr[*]}")"
-
-
-MODEL_ARGS=(
-   --spec "miles_plugins.models.qwen3_5" "get_qwen3_5_spec"
-
-   --disable-bias-linear
-   --qk-layernorm
-   --group-query-attention
-   --num-attention-heads 16
-   --num-query-groups 2
-   --kv-channels 256
-   --num-layers 40
-   --hidden-size 2048
-   --ffn-hidden-size 512
-
-   --normalization RMSNorm
-   --apply-layernorm-1p
-   --position-embedding-type rope
-   --norm-epsilon 1e-6
-   --rotary-percent 0.25
-   --swiglu
-   --untie-embeddings-and-output-weights
-   --vocab-size 248320
-
-   --rotary-base 10000000
-
-   # moe
-   --moe-ffn-hidden-size 512
-   --moe-shared-expert-intermediate-size 512
-   --moe-router-score-function softmax
-   --moe-token-dispatcher-type alltoall
-   --moe-router-topk 8
-   --moe-layer-freq "$MOE_LAYER_FREQ"
-   --num-experts 256
-   --moe-grouped-gemm
-   --moe-token-drop-policy probs
-   --moe-router-dtype fp32
-   --moe-permute-fusion
-   --moe-aux-loss-coeff 0
-
-   # qwen3.5 specific
-   --attention-output-gate
-   --moe-shared-expert-gate
-   --mtp-num-layers 1
-)
diff --git a/scripts/models/qwen3.5-35B-A3B_lora.py b/scripts/models/qwen3.5-35B-A3B_lora.py
new file mode 100644
index 00000000000..e63ef67ee33
--- /dev/null
+++ b/scripts/models/qwen3.5-35B-A3B_lora.py
@@ -0,0 +1,46 @@
+from model_args_utils import moe_layer_freq
+
+
+NLAYERS = 40
+FIRST_K_DENSE_REPLACE = 0
+
+
+def model_args() -> str:
+    return (
+        "--spec miles_plugins.models.qwen3_5 get_qwen3_5_spec "
+        "--disable-bias-linear "
+        "--qk-layernorm "
+        "--group-query-attention "
+        "--num-attention-heads 16 "
+        "--num-query-groups 2 "
+        "--kv-channels 256 "
+        "--num-layers 40 "
+        "--hidden-size 2048 "
+        "--ffn-hidden-size 512 "
+        "--normalization RMSNorm "
+        "--apply-layernorm-1p "
+        "--position-embedding-type rope "
+        "--norm-epsilon 1e-6 "
+        "--rotary-percent 0.25 "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--vocab-size 248320 "
+        "--rotary-base 10000000 "
+        # moe
+        "--moe-ffn-hidden-size 512 "
+        "--moe-shared-expert-intermediate-size 512 "
+        "--moe-router-score-function softmax "
+        "--moe-token-dispatcher-type alltoall "
+        "--moe-router-topk 8 "
+        f"--moe-layer-freq {moe_layer_freq(nlayers=NLAYERS, first_k_dense_replace=FIRST_K_DENSE_REPLACE)} "
+        "--num-experts 256 "
+        "--moe-grouped-gemm "
+        "--moe-token-drop-policy probs "
+        "--moe-router-dtype fp32 "
+        "--moe-permute-fusion "
+        "--moe-aux-loss-coeff 0 "
+        # qwen3.5 specific
+        "--attention-output-gate "
+        "--moe-shared-expert-gate "
+        "--mtp-num-layers 1 "
+    )
diff --git a/scripts/models/qwen3.5-35B-A3B_lora.sh b/scripts/models/qwen3.5-35B-A3B_lora.sh
deleted file mode 100644
index 9efa764fffc..00000000000
--- a/scripts/models/qwen3.5-35B-A3B_lora.sh
+++ /dev/null
@@ -1,62 +0,0 @@
-NLAYERS=40
-FIRST_K_DENSE_REPLACE=0
-
-arr=()
-for ((i=0; i<NLAYERS; i++)); do
-  if (( i < FIRST_K_DENSE_REPLACE )); then
-    arr+=(0)
-  else
-    arr+=(1)
-  fi
-done
-
-printf -v MOE_LAYER_FREQ "[%s]" "$(IFS=', '; echo "${arr[*]}")"
-
-
-MODEL_ARGS=(
-   --spec "miles_plugins.models.qwen3_5" "get_qwen3_5_spec"
-
-   --disable-bias-linear
-   --qk-layernorm
-   --group-query-attention
-   --num-attention-heads 16
-   --num-query-groups 2
-   --kv-channels 256
-   --num-layers 40
-   --hidden-size 2048
-   --ffn-hidden-size 512
-
-   --normalization RMSNorm
-   --apply-layernorm-1p
-   --position-embedding-type rope
-   --norm-epsilon 1e-6
-   --rotary-percent 0.25
-   --swiglu
-   --untie-embeddings-and-output-weights
-   --vocab-size 248320
-
-   --rotary-base 10000000
-
-   # moe
-   --moe-ffn-hidden-size 512
-   --moe-shared-expert-intermediate-size 512
-   --moe-router-score-function softmax
-   --moe-token-dispatcher-type alltoall
-   --moe-router-topk 8
-   --moe-layer-freq "$MOE_LAYER_FREQ"
-   --num-experts 256
-   --moe-grouped-gemm
-   --moe-token-drop-policy probs
-   --moe-router-dtype fp32
-   --moe-permute-fusion
-   --moe-aux-loss-coeff 0
-
-   # qwen3.5 specific
-   --attention-output-gate
-   --moe-shared-expert-gate
-   --mtp-num-layers 1
-)
-
-# LoRA registry for scripts/run_qwen3_5_35b_a3b_lora.py: MODEL_ARGS carries the architecture only
-# (identical to qwen3.5-35B-A3B.sh); every LoRA / run-mode flag lives in the runner,
-# which always wins (argparse last-occurrence). --spec above is inert under bridge LoRA.
diff --git a/scripts/models/qwen3.5-4B.py b/scripts/models/qwen3.5-4B.py
new file mode 100644
index 00000000000..44f28ffa6e4
--- /dev/null
+++ b/scripts/models/qwen3.5-4B.py
@@ -0,0 +1,23 @@
+def model_args() -> str:
+    return (
+        "--spec miles_plugins.models.qwen3_5 get_qwen3_5_spec "
+        "--disable-bias-linear "
+        "--qk-layernorm "
+        "--group-query-attention "
+        "--num-attention-heads 16 "
+        "--num-query-groups 4 "
+        "--kv-channels 256 "
+        "--num-layers 32 "
+        "--hidden-size 2560 "
+        "--ffn-hidden-size 9216 "
+        "--normalization RMSNorm "
+        "--apply-layernorm-1p "
+        "--position-embedding-type rope "
+        "--norm-epsilon 1e-6 "
+        "--rotary-percent 0.25 "
+        "--swiglu "
+        "--vocab-size 248320 "
+        "--rotary-base 10000000 "
+        # qwen3.5 specific
+        "--attention-output-gate "
+    )
diff --git a/scripts/models/qwen3.5-4B.sh b/scripts/models/qwen3.5-4B.sh
deleted file mode 100644
index 180ad79763b..00000000000
--- a/scripts/models/qwen3.5-4B.sh
+++ /dev/null
@@ -1,26 +0,0 @@
-MODEL_ARGS=(
-   --spec "miles_plugins.models.qwen3_5" "get_qwen3_5_spec"
-
-   --disable-bias-linear
-   --qk-layernorm
-   --group-query-attention
-   --num-attention-heads 16
-   --num-query-groups 4
-   --kv-channels 256
-   --num-layers 32
-   --hidden-size 2560
-   --ffn-hidden-size 9216
-
-   --normalization RMSNorm
-   --apply-layernorm-1p
-   --position-embedding-type rope
-   --norm-epsilon 1e-6
-   --rotary-percent 0.25
-   --swiglu
-   --vocab-size 248320
-
-   --rotary-base 10000000
-
-   # qwen3.5 specific
-   --attention-output-gate
-)
diff --git a/scripts/models/qwen3.5-9B.py b/scripts/models/qwen3.5-9B.py
new file mode 100644
index 00000000000..38f4e05ad60
--- /dev/null
+++ b/scripts/models/qwen3.5-9B.py
@@ -0,0 +1,24 @@
+def model_args() -> str:
+    return (
+        "--spec miles_plugins.models.qwen3_5 get_qwen3_5_spec "
+        "--disable-bias-linear "
+        "--qk-layernorm "
+        "--group-query-attention "
+        "--num-attention-heads 16 "
+        "--num-query-groups 4 "
+        "--kv-channels 256 "
+        "--num-layers 32 "
+        "--hidden-size 4096 "
+        "--ffn-hidden-size 12288 "
+        "--normalization RMSNorm "
+        "--apply-layernorm-1p "
+        "--position-embedding-type rope "
+        "--norm-epsilon 1e-6 "
+        "--rotary-percent 0.25 "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--vocab-size 248320 "
+        "--rotary-base 10000000 "
+        # qwen3.5 specific
+        "--attention-output-gate "
+    )
diff --git a/scripts/models/qwen3.5-9B.sh b/scripts/models/qwen3.5-9B.sh
deleted file mode 100644
index 433e730ae64..00000000000
--- a/scripts/models/qwen3.5-9B.sh
+++ /dev/null
@@ -1,27 +0,0 @@
-MODEL_ARGS=(
-   --spec "miles_plugins.models.qwen3_5" "get_qwen3_5_spec"
-
-   --disable-bias-linear
-   --qk-layernorm
-   --group-query-attention
-   --num-attention-heads 16
-   --num-query-groups 4
-   --kv-channels 256
-   --num-layers 32
-   --hidden-size 4096
-   --ffn-hidden-size 12288
-
-   --normalization RMSNorm
-   --apply-layernorm-1p
-   --position-embedding-type rope
-   --norm-epsilon 1e-6
-   --rotary-percent 0.25
-   --swiglu
-   --untie-embeddings-and-output-weights
-   --vocab-size 248320
-
-   --rotary-base 10000000
-
-   # qwen3.5 specific
-   --attention-output-gate
-)
diff --git a/scripts/models/qwen3.6-27B.py b/scripts/models/qwen3.6-27B.py
new file mode 100644
index 00000000000..3d7404b4b97
--- /dev/null
+++ b/scripts/models/qwen3.6-27B.py
@@ -0,0 +1,24 @@
+def model_args() -> str:
+    return (
+        "--spec miles_plugins.models.qwen3_5 get_qwen3_5_spec "
+        "--disable-bias-linear "
+        "--qk-layernorm "
+        "--group-query-attention "
+        "--num-attention-heads 24 "
+        "--num-query-groups 4 "
+        "--kv-channels 256 "
+        "--num-layers 64 "
+        "--hidden-size 5120 "
+        "--ffn-hidden-size 17408 "
+        "--normalization RMSNorm "
+        "--apply-layernorm-1p "
+        "--position-embedding-type rope "
+        "--norm-epsilon 1e-6 "
+        "--rotary-percent 0.25 "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--vocab-size 248320 "
+        "--rotary-base 10000000 "
+        # qwen3.5-family specific
+        "--attention-output-gate "
+    )
diff --git a/scripts/models/qwen3.6-27B.sh b/scripts/models/qwen3.6-27B.sh
deleted file mode 100644
index c30e566d0a3..00000000000
--- a/scripts/models/qwen3.6-27B.sh
+++ /dev/null
@@ -1,27 +0,0 @@
-MODEL_ARGS=(
-   --spec "miles_plugins.models.qwen3_5" "get_qwen3_5_spec"
-
-   --disable-bias-linear
-   --qk-layernorm
-   --group-query-attention
-   --num-attention-heads 24
-   --num-query-groups 4
-   --kv-channels 256
-   --num-layers 64
-   --hidden-size 5120
-   --ffn-hidden-size 17408
-
-   --normalization RMSNorm
-   --apply-layernorm-1p
-   --position-embedding-type rope
-   --norm-epsilon 1e-6
-   --rotary-percent 0.25
-   --swiglu
-   --untie-embeddings-and-output-weights
-   --vocab-size 248320
-
-   --rotary-base 10000000
-
-   # qwen3.5-family specific
-   --attention-output-gate
-)
diff --git a/scripts/models/qwen3.6-35B-A3B.py b/scripts/models/qwen3.6-35B-A3B.py
new file mode 100644
index 00000000000..6dea958668d
--- /dev/null
+++ b/scripts/models/qwen3.6-35B-A3B.py
@@ -0,0 +1,46 @@
+from model_args_utils import moe_layer_freq
+
+
+NLAYERS = 40
+FIRST_K_DENSE_REPLACE = 0
+
+
+def model_args() -> str:
+    return (
+        "--spec miles_plugins.models.qwen3_5 get_qwen3_5_spec "
+        "--disable-bias-linear "
+        "--qk-layernorm "
+        "--group-query-attention "
+        "--num-attention-heads 16 "
+        "--num-query-groups 2 "
+        "--kv-channels 256 "
+        "--num-layers 40 "
+        "--hidden-size 2048 "
+        "--ffn-hidden-size 512 "
+        "--normalization RMSNorm "
+        "--apply-layernorm-1p "
+        "--position-embedding-type rope "
+        "--norm-epsilon 1e-6 "
+        "--rotary-percent 0.25 "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--vocab-size 248320 "
+        "--rotary-base 10000000 "
+        # moe
+        "--moe-ffn-hidden-size 512 "
+        "--moe-shared-expert-intermediate-size 512 "
+        "--moe-router-score-function softmax "
+        "--moe-token-dispatcher-type alltoall "
+        "--moe-router-topk 8 "
+        f"--moe-layer-freq {moe_layer_freq(nlayers=NLAYERS, first_k_dense_replace=FIRST_K_DENSE_REPLACE)} "
+        "--num-experts 256 "
+        "--moe-grouped-gemm "
+        "--moe-token-drop-policy probs "
+        "--moe-router-dtype fp32 "
+        "--moe-permute-fusion "
+        "--moe-aux-loss-coeff 0 "
+        # qwen3.6 specific (same architecture as qwen3.5)
+        "--attention-output-gate "
+        "--moe-shared-expert-gate "
+        "--mtp-num-layers 1 "
+    )
diff --git a/scripts/models/qwen3.6-35B-A3B.sh b/scripts/models/qwen3.6-35B-A3B.sh
deleted file mode 100644
index 48d323213a6..00000000000
--- a/scripts/models/qwen3.6-35B-A3B.sh
+++ /dev/null
@@ -1,58 +0,0 @@
-NLAYERS=40
-FIRST_K_DENSE_REPLACE=0
-
-arr=()
-for ((i=0; i<NLAYERS; i++)); do
-  if (( i < FIRST_K_DENSE_REPLACE )); then
-    arr+=(0)
-  else
-    arr+=(1)
-  fi
-done
-
-printf -v MOE_LAYER_FREQ "[%s]" "$(IFS=', '; echo "${arr[*]}")"
-
-
-MODEL_ARGS=(
-   --spec "miles_plugins.models.qwen3_5" "get_qwen3_5_spec"
-
-   --disable-bias-linear
-   --qk-layernorm
-   --group-query-attention
-   --num-attention-heads 16
-   --num-query-groups 2
-   --kv-channels 256
-   --num-layers 40
-   --hidden-size 2048
-   --ffn-hidden-size 512
-
-   --normalization RMSNorm
-   --apply-layernorm-1p
-   --position-embedding-type rope
-   --norm-epsilon 1e-6
-   --rotary-percent 0.25
-   --swiglu
-   --untie-embeddings-and-output-weights
-   --vocab-size 248320
-
-   --rotary-base 10000000
-
-   # moe
-   --moe-ffn-hidden-size 512
-   --moe-shared-expert-intermediate-size 512
-   --moe-router-score-function softmax
-   --moe-token-dispatcher-type alltoall
-   --moe-router-topk 8
-   --moe-layer-freq "$MOE_LAYER_FREQ"
-   --num-experts 256
-   --moe-grouped-gemm
-   --moe-token-drop-policy probs
-   --moe-router-dtype fp32
-   --moe-permute-fusion
-   --moe-aux-loss-coeff 0
-
-   # qwen3.6 specific (same architecture as qwen3.5)
-   --attention-output-gate
-   --moe-shared-expert-gate
-   --mtp-num-layers 1
-)
diff --git a/scripts/models/qwen3.6-35B-A3B_lora.py b/scripts/models/qwen3.6-35B-A3B_lora.py
new file mode 100644
index 00000000000..6dea958668d
--- /dev/null
+++ b/scripts/models/qwen3.6-35B-A3B_lora.py
@@ -0,0 +1,46 @@
+from model_args_utils import moe_layer_freq
+
+
+NLAYERS = 40
+FIRST_K_DENSE_REPLACE = 0
+
+
+def model_args() -> str:
+    return (
+        "--spec miles_plugins.models.qwen3_5 get_qwen3_5_spec "
+        "--disable-bias-linear "
+        "--qk-layernorm "
+        "--group-query-attention "
+        "--num-attention-heads 16 "
+        "--num-query-groups 2 "
+        "--kv-channels 256 "
+        "--num-layers 40 "
+        "--hidden-size 2048 "
+        "--ffn-hidden-size 512 "
+        "--normalization RMSNorm "
+        "--apply-layernorm-1p "
+        "--position-embedding-type rope "
+        "--norm-epsilon 1e-6 "
+        "--rotary-percent 0.25 "
+        "--swiglu "
+        "--untie-embeddings-and-output-weights "
+        "--vocab-size 248320 "
+        "--rotary-base 10000000 "
+        # moe
+        "--moe-ffn-hidden-size 512 "
+        "--moe-shared-expert-intermediate-size 512 "
+        "--moe-router-score-function softmax "
+        "--moe-token-dispatcher-type alltoall "
+        "--moe-router-topk 8 "
+        f"--moe-layer-freq {moe_layer_freq(nlayers=NLAYERS, first_k_dense_replace=FIRST_K_DENSE_REPLACE)} "
+        "--num-experts 256 "
+        "--moe-grouped-gemm "
+        "--moe-token-drop-policy probs "
+        "--moe-router-dtype fp32 "
+        "--moe-permute-fusion "
+        "--moe-aux-loss-coeff 0 "
+        # qwen3.6 specific (same architecture as qwen3.5)
+        "--attention-output-gate "
+        "--moe-shared-expert-gate "
+        "--mtp-num-layers 1 "
+    )
diff --git a/scripts/models/qwen3.6-35B-A3B_lora.sh b/scripts/models/qwen3.6-35B-A3B_lora.sh
deleted file mode 100644
index cca1ee11399..00000000000
--- a/scripts/models/qwen3.6-35B-A3B_lora.sh
+++ /dev/null
@@ -1,62 +0,0 @@
-NLAYERS=40
-FIRST_K_DENSE_REPLACE=0
-
-arr=()
-for ((i=0; i<NLAYERS; i++)); do
-  if (( i < FIRST_K_DENSE_REPLACE )); then
-    arr+=(0)
-  else
-    arr+=(1)
-  fi
-done
-
-printf -v MOE_LAYER_FREQ "[%s]" "$(IFS=', '; echo "${arr[*]}")"
-
-
-MODEL_ARGS=(
-   --spec "miles_plugins.models.qwen3_5" "get_qwen3_5_spec"
-
-   --disable-bias-linear
-   --qk-layernorm
-   --group-query-attention
-   --num-attention-heads 16
-   --num-query-groups 2
-   --kv-channels 256
-   --num-layers 40
-   --hidden-size 2048
-   --ffn-hidden-size 512
-
-   --normalization RMSNorm
-   --apply-layernorm-1p
-   --position-embedding-type rope
-   --norm-epsilon 1e-6
-   --rotary-percent 0.25
-   --swiglu
-   --untie-embeddings-and-output-weights
-   --vocab-size 248320
-
-   --rotary-base 10000000
-
-   # moe
-   --moe-ffn-hidden-size 512
-   --moe-shared-expert-intermediate-size 512
-   --moe-router-score-function softmax
-   --moe-token-dispatcher-type alltoall
-   --moe-router-topk 8
-   --moe-layer-freq "$MOE_LAYER_FREQ"
-   --num-experts 256
-   --moe-grouped-gemm
-   --moe-token-drop-policy probs
-   --moe-router-dtype fp32
-   --moe-permute-fusion
-   --moe-aux-loss-coeff 0
-
-   # qwen3.6 specific (same architecture as qwen3.5)
-   --attention-output-gate
-   --moe-shared-expert-gate
-   --mtp-num-layers 1
-)
-
-# Qwen3.6 (same architecture as qwen3.5). LoRA registry for scripts/run_qwen3_5_35b_a3b_lora.py: MODEL_ARGS carries the architecture only
-# (identical to qwen3.5-35B-A3B.sh); every LoRA / run-mode flag lives in the runner,
-# which always wins (argparse last-occurrence). --spec above is inert under bridge LoRA.
diff --git a/scripts/run-deepseek-r1.sh b/scripts/run-deepseek-r1.sh
index 35e3d313e30..26a941c79ff 100644
--- a/scripts/run-deepseek-r1.sh
+++ b/scripts/run-deepseek-r1.sh
@@ -24,8 +24,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/deepseek-v3.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "deepseek-v3")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint $BASE_DIR/DeepSeek-R1/
    #--hf-checkpoint $BASE_DIR/DeepSeek-R1-bf16/
diff --git a/scripts/run-glm4-9B-4xgpu-radixtree.sh b/scripts/run-glm4-9B-4xgpu-radixtree.sh
index dbebcd3782b..2d0b0f2a179 100755
--- a/scripts/run-glm4-9B-4xgpu-radixtree.sh
+++ b/scripts/run-glm4-9B-4xgpu-radixtree.sh
@@ -26,8 +26,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/glm4-9B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "glm4-9B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint /root/GLM-Z1-9B-0414/
    --ref-load /root/GLM-Z1-9B-0414_torch_dist
diff --git a/scripts/run-glm4-9B.sh b/scripts/run-glm4-9B.sh
index 84080ae63bb..b45ecde9885 100644
--- a/scripts/run-glm4-9B.sh
+++ b/scripts/run-glm4-9B.sh
@@ -24,8 +24,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/glm4-9B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "glm4-9B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint /root/GLM-Z1-9B-0414/
    --ref-load /root/GLM-Z1-9B-0414_torch_dist
diff --git a/scripts/run-glm4.5-355B-A32B.sh b/scripts/run-glm4.5-355B-A32B.sh
index 36e3366e0ca..dc0c1a69441 100644
--- a/scripts/run-glm4.5-355B-A32B.sh
+++ b/scripts/run-glm4.5-355B-A32B.sh
@@ -24,8 +24,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/glm4.5-355B-A32B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "glm4.5-355B-A32B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint $BASE_DIR/GLM-4.5-355B-A32B
    --ref-load $BASE_DIR/GLM-4.5-355B-A32B_torch_dist/
diff --git a/scripts/run-glm4.7-flash.sh b/scripts/run-glm4.7-flash.sh
index 18e58fa88e1..06a8192fad4 100644
--- a/scripts/run-glm4.7-flash.sh
+++ b/scripts/run-glm4.7-flash.sh
@@ -24,8 +24,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/../scripts/models/glm4.7-flash.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "glm4.7-flash")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 BASE_DIR=/root/shared
 
 CKPT_ARGS=(
diff --git a/scripts/run-gpt-oss-20b-bf16.sh b/scripts/run-gpt-oss-20b-bf16.sh
index 6ad71ce4c26..4cf823ee1b4 100644
--- a/scripts/run-gpt-oss-20b-bf16.sh
+++ b/scripts/run-gpt-oss-20b-bf16.sh
@@ -18,8 +18,8 @@ export HF_HOME=/workspace/hf_cache
 
 # Load model architecture config
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/gpt-oss-20b.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "gpt-oss-20b")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 BASE_DIR=/root/shared
 
 CKPT_ARGS=(
diff --git a/scripts/run-kimi-k2-Instruct.sh b/scripts/run-kimi-k2-Instruct.sh
index 525f63c6a71..f2e64742034 100644
--- a/scripts/run-kimi-k2-Instruct.sh
+++ b/scripts/run-kimi-k2-Instruct.sh
@@ -24,8 +24,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/kimi-k2.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "kimi-k2")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint $BASE_DIR/Kimi-K2-Instruct/
    # --hf-checkpoint $BASE_DIR/Kimi-K2-bf16/
diff --git a/scripts/run-kimi-k2-Thinking.sh b/scripts/run-kimi-k2-Thinking.sh
index d603fedb472..da6b0371e15 100644
--- a/scripts/run-kimi-k2-Thinking.sh
+++ b/scripts/run-kimi-k2-Thinking.sh
@@ -24,8 +24,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/kimi-k2-thinking.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "kimi-k2-thinking")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(   
    # --hf-checkpoint $BASE_DIR/Kimi-K2-Thinking-bf16/
    --hf-checkpoint $BASE_DIR/Kimi-K2-Thinking-fp8/
diff --git a/scripts/run-kimi-k25.sh b/scripts/run-kimi-k25.sh
index e0ec3cccccc..01c82e7e1c4 100755
--- a/scripts/run-kimi-k25.sh
+++ b/scripts/run-kimi-k25.sh
@@ -27,8 +27,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/kimi-k2-thinking.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "kimi-k2-thinking")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint $BASE_DIR/Kimi-K2.5-int4
    --ref-load $BASE_DIR/Kimi-K2.5-bf16
diff --git a/scripts/run-mimo-7B-rl-eagle.sh b/scripts/run-mimo-7B-rl-eagle.sh
index 2efbc2d6be8..3d7fa6f7711 100644
--- a/scripts/run-mimo-7B-rl-eagle.sh
+++ b/scripts/run-mimo-7B-rl-eagle.sh
@@ -25,8 +25,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/mimo-7B-rl.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "mimo-7B-rl")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
  CKPT_ARGS=(
    --hf-checkpoint /root/MiMo-7B-RL
    #--hf-checkpoint /root/Qwen3-4B-FP8
diff --git a/scripts/run-moonlight-16B-A3B.sh b/scripts/run-moonlight-16B-A3B.sh
index 69a66fdfc4d..772a7c279f0 100644
--- a/scripts/run-moonlight-16B-A3B.sh
+++ b/scripts/run-moonlight-16B-A3B.sh
@@ -25,8 +25,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/moonlight.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "moonlight")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint /root/Moonlight-16B-A3B
    --ref-load /root/Moonlight-16B-A3B_torch_dist
diff --git a/scripts/run-nemotron-3-nano-30b-a3b.sh b/scripts/run-nemotron-3-nano-30b-a3b.sh
index 4123adb8820..c405eead797 100755
--- a/scripts/run-nemotron-3-nano-30b-a3b.sh
+++ b/scripts/run-nemotron-3-nano-30b-a3b.sh
@@ -21,8 +21,8 @@ if [ "$NVLINK_COUNT" -gt 0 ]; then HAS_NVLINK=1; else HAS_NVLINK=0; fi
 echo "HAS_NVLINK: $HAS_NVLINK"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/nemotron-3-nano-30b-a3b.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "nemotron-3-nano-30b-a3b")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint $BASE_DIR/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
    --ref-load $BASE_DIR/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
diff --git a/scripts/run-nemotron-3-nano-4b.sh b/scripts/run-nemotron-3-nano-4b.sh
index dfdfe7743b6..3316085ba6c 100644
--- a/scripts/run-nemotron-3-nano-4b.sh
+++ b/scripts/run-nemotron-3-nano-4b.sh
@@ -26,8 +26,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/nemotron-3-nano-4b.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "nemotron-3-nano-4b")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint $BASE_DIR/NVIDIA-Nemotron-3-Nano-4B-BF16
    --ref-load $BASE_DIR/NVIDIA-Nemotron-3-Nano-4B-BF16
diff --git a/scripts/run-nemotron-3-super-120b-a12b.sh b/scripts/run-nemotron-3-super-120b-a12b.sh
index fd1e31f1461..54f77e66b84 100755
--- a/scripts/run-nemotron-3-super-120b-a12b.sh
+++ b/scripts/run-nemotron-3-super-120b-a12b.sh
@@ -45,8 +45,8 @@ if [[ "$ROLE" == "worker" ]]; then
 fi
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/nemotron-3-super-120b-a12b.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "nemotron-3-super-120b-a12b")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 MODELS_DIR=${MODELS_DIR:-/cluster_public/miles_data/models}
 DATASETS_DIR=${DATASETS_DIR:-/cluster_public/miles_data/datasets}
 
diff --git a/scripts/run-qwen3-235B-A22B-sft.sh b/scripts/run-qwen3-235B-A22B-sft.sh
index a5a801c4c8f..3233241189c 100644
--- a/scripts/run-qwen3-235B-A22B-sft.sh
+++ b/scripts/run-qwen3-235B-A22B-sft.sh
@@ -35,8 +35,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/qwen3-235B-A22B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "qwen3-235B-A22B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint ${BASE_FOLDER}/Qwen3-235B-A22B
    --ref-load ${BASE_FOLDER}/Qwen3-235B-A22B_torch_dist
diff --git a/scripts/run-qwen3-235B-A22B.sh b/scripts/run-qwen3-235B-A22B.sh
index 45067036fb2..51cb60afbd0 100644
--- a/scripts/run-qwen3-235B-A22B.sh
+++ b/scripts/run-qwen3-235B-A22B.sh
@@ -35,8 +35,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/qwen3-235B-A22B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "qwen3-235B-A22B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint ${BASE_FOLDER}/Qwen3-235B-A22B-FP8
    --ref-load ${BASE_FOLDER}/Qwen3-235B-A22B_torch_dist
diff --git a/scripts/run-qwen3-32B.sh b/scripts/run-qwen3-32B.sh
index 156bcf5d03a..92b5f6ce30f 100644
--- a/scripts/run-qwen3-32B.sh
+++ b/scripts/run-qwen3-32B.sh
@@ -24,8 +24,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/qwen3-32B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "qwen3-32B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint /root/Qwen3-32B
    --ref-load /root/Qwen3-32B_torch_dist/
diff --git a/scripts/run-qwen3-4B-base-sft.sh b/scripts/run-qwen3-4B-base-sft.sh
index a30209f750f..e8acefd9ea4 100644
--- a/scripts/run-qwen3-4B-base-sft.sh
+++ b/scripts/run-qwen3-4B-base-sft.sh
@@ -24,8 +24,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/qwen3-4B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "qwen3-4B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint /root/Qwen3-4B-Base/
    --ref-load /root/Qwen3-4B-Base_torch_dist
diff --git a/scripts/run-qwen3-4B.sh b/scripts/run-qwen3-4B.sh
index 2285cf57c0a..d52e732fbe2 100644
--- a/scripts/run-qwen3-4B.sh
+++ b/scripts/run-qwen3-4B.sh
@@ -24,8 +24,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/qwen3-4B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "qwen3-4B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint /root/Qwen3-4B
    #--hf-checkpoint /root/Qwen3-4B-FP8
diff --git a/scripts/run-qwen3-4B_4xgpu.sh b/scripts/run-qwen3-4B_4xgpu.sh
index 085266c7dd8..54a268e657b 100755
--- a/scripts/run-qwen3-4B_4xgpu.sh
+++ b/scripts/run-qwen3-4B_4xgpu.sh
@@ -26,8 +26,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/qwen3-4B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "qwen3-4B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint /root/Qwen3-4B
    #--hf-checkpoint /root/Qwen3-4B-FP8
diff --git a/scripts/run-qwen3-next-80B-A3B-8gpus.sh b/scripts/run-qwen3-next-80B-A3B-8gpus.sh
index bc110cf43b7..e8e2aee2c8d 100644
--- a/scripts/run-qwen3-next-80B-A3B-8gpus.sh
+++ b/scripts/run-qwen3-next-80B-A3B-8gpus.sh
@@ -35,8 +35,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/qwen3-next-80B-A3B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "qwen3-next-80B-A3B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint ${BASE_FOLDER}/Qwen3-Next-80B-A3B-Thinking
    --ref-load ${BASE_FOLDER}/Qwen3-Next-80B-A3B-Thinking_torch_dist
diff --git a/scripts/run-qwen3-next-80B-A3B.sh b/scripts/run-qwen3-next-80B-A3B.sh
index 545c8a1309a..6973de821f2 100644
--- a/scripts/run-qwen3-next-80B-A3B.sh
+++ b/scripts/run-qwen3-next-80B-A3B.sh
@@ -35,8 +35,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/qwen3-next-80B-A3B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "qwen3-next-80B-A3B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint ${BASE_FOLDER}/Qwen3-Next-80B-A3B-Thinking
    --ref-load ${BASE_FOLDER}/Qwen3-Next-80B-A3B-Thinking_torch_dist
diff --git a/scripts/run-qwen3.5-27B.sh b/scripts/run-qwen3.5-27B.sh
index 3eab260fb18..4221e996caa 100644
--- a/scripts/run-qwen3.5-27B.sh
+++ b/scripts/run-qwen3.5-27B.sh
@@ -24,8 +24,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/qwen3.5-27B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "qwen3.5-27B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint /root/Qwen3.5-27B
    --ref-load /root/Qwen3.5-27B_torch_dist
diff --git a/scripts/run-qwen3.5-35B-A3B-mtp.sh b/scripts/run-qwen3.5-35B-A3B-mtp.sh
index 062d99686e5..cc8122645ad 100755
--- a/scripts/run-qwen3.5-35B-A3B-mtp.sh
+++ b/scripts/run-qwen3.5-35B-A3B-mtp.sh
@@ -24,8 +24,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/qwen3.5-35B-A3B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "qwen3.5-35B-A3B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint /root/Qwen3.5-35B-A3B
    --ref-load /root/Qwen3.5-35B-A3B_torch_dist
diff --git a/scripts/run-qwen3.5-4B.sh b/scripts/run-qwen3.5-4B.sh
index 7fce9bdae9e..c9278f6ef8f 100644
--- a/scripts/run-qwen3.5-4B.sh
+++ b/scripts/run-qwen3.5-4B.sh
@@ -24,8 +24,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/qwen3.5-4B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "qwen3.5-4B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint /root/Qwen3.5-4B
    --ref-load /root/Qwen3.5-4B_torch_dist
diff --git a/scripts/run-qwen3.5-9B.sh b/scripts/run-qwen3.5-9B.sh
index 7664feb47b8..66db36dfaf4 100644
--- a/scripts/run-qwen3.5-9B.sh
+++ b/scripts/run-qwen3.5-9B.sh
@@ -24,8 +24,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/qwen3.5-9B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "qwen3.5-9B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CKPT_ARGS=(
    --hf-checkpoint /root/Qwen3.5-9B
    --ref-load /root/Qwen3.5-9B_torch_dist
diff --git a/scripts/run-qwen3.6-27B.sh b/scripts/run-qwen3.6-27B.sh
index dbe7c7c2c07..c7b66f2ad12 100755
--- a/scripts/run-qwen3.6-27B.sh
+++ b/scripts/run-qwen3.6-27B.sh
@@ -24,8 +24,8 @@ fi
 echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
 
 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
-source "${SCRIPT_DIR}/models/qwen3.6-27B.sh"
-
+MODEL_ARGS_LINE="$(python3 "${SCRIPT_DIR}/../miles/utils/external_utils/model_args_utils.py" "qwen3.6-27B")" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 MODEL_DIR="${MODEL_DIR:-/cluster_public/miles_data/models}"
 DATA_DIR="${DATA_DIR:-/cluster_public/miles_data/datasets}"
 OUTPUT_DIR="${OUTPUT_DIR:?set OUTPUT_DIR to a writable checkpoint directory}"
diff --git a/scripts/run_gemma_4_26b_a4b.py b/scripts/run_gemma_4_26b_a4b.py
index f654badc13e..c87b86438ad 100644
--- a/scripts/run_gemma_4_26b_a4b.py
+++ b/scripts/run_gemma_4_26b_a4b.py
@@ -4,7 +4,7 @@
 Trained via the HF<->Megatron bridge (`--megatron-to-hf-mode bridge`) on the base
 VLM checkpoint directly — sglang runs Gemma4ForConditionalGeneration (hybrid swa),
 which loads gemma-4's hybrid head_dim weights correctly. MODEL_ARGS come from
-scripts/models/gemma-4-26b-a4b-it.sh.
+scripts/models/gemma-4-26b-a4b-it.py.
 
 Single-node smoke test:
   python scripts/run_gemma_4_26b_a4b.py full-train --num-nodes 1
diff --git a/scripts/run_gemma_4_31b.py b/scripts/run_gemma_4_31b.py
index 8e8f40a67a1..81d83739de1 100644
--- a/scripts/run_gemma_4_31b.py
+++ b/scripts/run_gemma_4_31b.py
@@ -5,7 +5,7 @@
 parallelism. Trained via the HF<->Megatron bridge (`--megatron-to-hf-mode
 bridge`); the dense config is driven directly through Gemma4VLBridge, so there
 is no LLM-view rewrite or offline conversion — `prepare` only downloads.
-MODEL_ARGS come from scripts/models/gemma-4-31b-it.sh.
+MODEL_ARGS come from scripts/models/gemma-4-31b-it.py.
 
 Requires the radixark/Megatron-Bridge gemma4-dense branch.
 
diff --git a/scripts/run_inkling.py b/scripts/run_inkling.py
index df3c59e8d0b..a59a15952c2 100644
--- a/scripts/run_inkling.py
+++ b/scripts/run_inkling.py
@@ -68,7 +68,7 @@
 
 app = typer.Typer()
 
-# model name -> scripts/models/<type>.sh; the 4-layer slices reuse the base
+# model name -> scripts/models/<type>.py; the 4-layer slices reuse the base
 # definition with MODEL_ARGS_NUM_LAYERS=4 (set in ScriptArgs.__post_init__)
 _MODEL_REGISTRY = {
     "Inkling": "inkling",
diff --git a/scripts/run_kimi_k25.py b/scripts/run_kimi_k25.py
index 589c540f20b..00e92821c33 100644
--- a/scripts/run_kimi_k25.py
+++ b/scripts/run_kimi_k25.py
@@ -8,7 +8,7 @@
 weights for the SGLang rollout while Megatron loads a BF16 reference via the
 HF<->Megatron bridge (`--megatron-to-hf-mode bridge`), so there is no offline
 `torch_dist` conversion step. The architecture is shared with Kimi-K2-Thinking,
-whose Megatron MODEL_ARGS we reuse (`scripts/models/kimi-k2-thinking.sh`).
+whose Megatron MODEL_ARGS we reuse (`scripts/models/kimi-k2-thinking.py`).
 
 =====================
 
diff --git a/tests/e2e/sglang/test_r3_router_equivalence.py b/tests/e2e/sglang/test_r3_router_equivalence.py
index 5ade9d0d43a..3334cdfe656 100644
--- a/tests/e2e/sglang/test_r3_router_equivalence.py
+++ b/tests/e2e/sglang/test_r3_router_equivalence.py
@@ -32,7 +32,7 @@
 Backend / checkpoint
 ~~~~~~~~~~~~~~~~~~~~
 Megatron backend (same as the sibling ``tests/e2e/megatron/*_r3.py``
-tests) — sourcing ``scripts/models/{type}.sh`` populates
+tests) — loading ``scripts/models/{type}.py`` populates
 ``args.num_layers`` / ``args.moe_router_topk`` that the rollout-side
 reshape of ``routed_experts`` depends on.  We do *not* set
 ``--use-kl-loss`` or ``--kl-coef`` > 0, which is what gates the
diff --git a/tests/fast/launch_scripts/model_args_harness.py b/tests/fast/launch_scripts/model_args_harness.py
index ceaa42a47d3..516927958e9 100644
--- a/tests/fast/launch_scripts/model_args_harness.py
+++ b/tests/fast/launch_scripts/model_args_harness.py
@@ -1,31 +1,14 @@
-import subprocess
-
 from tests.fast.launch_scripts.sh_harness import REPO_ROOT
 
-MODEL_SCRIPT_DIR = REPO_ROOT / "scripts" / "models"
+from miles.utils.external_utils.model_args_utils import load_model_args
 
-_ENV_WITHOUT_THE_MODEL_ARGS_KNOBS = {
-    "PATH": "/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin",
-    "HOME": "/root",
-    "LANG": "C",
-    "LC_ALL": "C",
-}
+MODEL_SCRIPT_DIR = REPO_ROOT / "scripts" / "models"
 
 
 def iter_model_types() -> list[str]:
-    return sorted(path.stem for path in MODEL_SCRIPT_DIR.glob("*.sh"))
+    return sorted(path.stem for path in MODEL_SCRIPT_DIR.glob("*.py"))
 
 
 def expand_model_args(model_type: str) -> list[str]:
-    """The golden files are taken from this shell expansion; whatever replaces it must reproduce them."""
-    script = MODEL_SCRIPT_DIR / f"{model_type}.sh"
-    result = subprocess.run(
-        f'source "{script}" && printf "%s\\n" "${{MODEL_ARGS[@]}}"',
-        shell=True,
-        executable="/bin/bash",
-        env=_ENV_WITHOUT_THE_MODEL_ARGS_KNOBS,
-        capture_output=True,
-        text=True,
-        check=True,
-    )
-    return result.stdout.splitlines()
+    """Only the producer changed here; the golden files still hold what the shell era expanded to."""
+    return load_model_args(model_type).split()
diff --git a/tests/fast/launch_scripts/py_harness.py b/tests/fast/launch_scripts/py_harness.py
index d50baf9e67d..9690989669c 100644
--- a/tests/fast/launch_scripts/py_harness.py
+++ b/tests/fast/launch_scripts/py_harness.py
@@ -1,10 +1,8 @@
 import ast
-import importlib.util
 import inspect
 import os
 import re
 import subprocess
-import sys
 import time
 from collections.abc import Iterator
 from contextlib import contextmanager
@@ -16,6 +14,7 @@
 from tests.fast.utils.command_recorder import record_commands
 
 import miles.utils.external_utils.command_utils as command_utils
+from miles.utils.external_utils.model_args_utils import import_module_from_path
 
 FROZEN_RUN_ID = "260101-000000-000"
 
@@ -126,15 +125,7 @@ def fake_encode_pseudo_file(text: str) -> str:
 
 def import_launch_script(path: Path) -> ModuleType:
     name = "miles_launch_script_" + path.relative_to(REPO_ROOT).with_suffix("").as_posix().replace("/", "_")
-    spec = importlib.util.spec_from_file_location(name, path)
-    assert spec is not None and spec.loader is not None
-    module = importlib.util.module_from_spec(spec)
-    sys.modules[name] = module
-    try:
-        spec.loader.exec_module(module)
-    finally:
-        del sys.modules[name]
-    return module
+    return import_module_from_path(path, name)
 
 
 @contextmanager
diff --git a/tests/fast/launch_scripts/sh_harness.py b/tests/fast/launch_scripts/sh_harness.py
index c3069e8f70d..16bbbfbab32 100644
--- a/tests/fast/launch_scripts/sh_harness.py
+++ b/tests/fast/launch_scripts/sh_harness.py
@@ -66,6 +66,7 @@
 }
 
 _PYTHON_SHIM_BODY = """case "${1:-}" in
+*/model_args_utils.py) "$MILES_SH_HARNESS_REAL_PYTHON" "$@" ;;
 -c)
     case "$2" in
     *cluster_resources*) printf '%s\\n' 'REPLACE_GPU_COUNT' ;;
diff --git a/tests/fast/launch_scripts/test_sh_harness.py b/tests/fast/launch_scripts/test_sh_harness.py
index c0de9fc1b07..f91a3163b7c 100644
--- a/tests/fast/launch_scripts/test_sh_harness.py
+++ b/tests/fast/launch_scripts/test_sh_harness.py
@@ -60,7 +60,7 @@ def test_ray_start_is_recorded_with_the_frozen_master_addr(self, run):
         assert ray_start[ray_start.index("--node-ip-address") + 1] == "127.0.0.1"
 
     def test_ray_job_submit_argv_contains_the_expanded_model_args(self, run):
-        """`source scripts/models/*.sh` expansion must be visible in the captured argv."""
+        """The scripts/models/*.py expansion must be visible in the captured argv."""
         argv = run.ray_job_submit_argv()
         assert argv[:3] == ["ray", "job", "submit"]
         assert "--num-layers" in argv
diff --git a/tests/fast/launch_scripts/test_shell_script_hygiene.py b/tests/fast/launch_scripts/test_shell_script_hygiene.py
index 22c4502e75f..501012d41d4 100644
--- a/tests/fast/launch_scripts/test_shell_script_hygiene.py
+++ b/tests/fast/launch_scripts/test_shell_script_hygiene.py
@@ -6,6 +6,10 @@
 
 _REMOVED_COMMAND_HELPERS = re.compile(r"(?<![\w.])exec_command\s*\(")
 
+_REMOVED_MODEL_SCRIPTS = re.compile(r"scripts/models/\S+\.sh")
+
+_JOINED_MODEL_ARGS = re.compile(r"join\(\s*load_model_args")
+
 
 class TestShellScriptHygiene:
     def test_no_shell_script_hardcodes_the_checkout_location(self):
@@ -31,3 +35,23 @@ def test_no_patch_calls_the_removed_exec_command_helper(self):
         ]
 
         assert offenders == []
+
+    def test_no_patch_joins_the_model_args_line(self):
+        """load_model_args() returns the line; joining it would splice the string character by character."""
+        offenders = [
+            path.relative_to(REPO_ROOT).as_posix()
+            for path in (REPO_ROOT / "docker").rglob("*.patch")
+            if _JOINED_MODEL_ARGS.search(path.read_text(errors="replace"))
+        ]
+
+        assert offenders == []
+
+    def test_no_patch_sources_a_removed_shell_model_script(self):
+        """The model definitions are python now, so a `source scripts/models/x.sh` can never resolve."""
+        offenders = [
+            path.relative_to(REPO_ROOT).as_posix()
+            for path in (REPO_ROOT / "docker").rglob("*.patch")
+            if _REMOVED_MODEL_SCRIPTS.search(path.read_text(errors="replace"))
+        ]
+
+        assert offenders == []
diff --git a/tests/fast/utils/debug_utils/run_megatron/cli/commands/test_run.py b/tests/fast/utils/debug_utils/run_megatron/cli/commands/test_run.py
index f7651464c48..587287d4de0 100644
--- a/tests/fast/utils/debug_utils/run_megatron/cli/commands/test_run.py
+++ b/tests/fast/utils/debug_utils/run_megatron/cli/commands/test_run.py
@@ -42,7 +42,7 @@ def test_routing_replay_pp_gt1_also_fails(self) -> None:
 
 class TestRunImplExecCommand:
     """Only mock exec_command_gpu, generate_token_ids, write_token_ids_to_tmpfile,
-    and resolve_model_script — let the rest (build_worker_args, build_dumper_env,
+    and load_model_args — let the rest (build_worker_args, build_dumper_env,
     build_torchrun_cmd, ParallelConfig, WorkerScriptArgs) run for real."""
 
     @pytest.fixture(autouse=True)
@@ -58,8 +58,8 @@ def _patch_externals(self) -> Generator[None, None, None]:
                 return_value=Path("/tmp/tokens.json"),
             ),
             patch(
-                "miles.utils.debug_utils.run_megatron.cli.worker_executor.resolve_model_script",
-                return_value=Path("/repo/scripts/models/deepseek_v3.sh"),
+                "miles.utils.debug_utils.run_megatron.cli.worker_executor.load_model_args",
+                return_value="--num-layers 61",
             ),
         ):
             self.mock_exec = mock_exec
diff --git a/tests/fast/utils/debug_utils/run_megatron/cli/test_path_utils.py b/tests/fast/utils/debug_utils/run_megatron/cli/test_path_utils.py
index 3858d58d24a..7c9790ec972 100644
--- a/tests/fast/utils/debug_utils/run_megatron/cli/test_path_utils.py
+++ b/tests/fast/utils/debug_utils/run_megatron/cli/test_path_utils.py
@@ -29,7 +29,7 @@ class TestResolveModelScript:
     def test_returns_path_when_exists(self, tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
         scripts_dir = tmp_path / "scripts" / "models"
         scripts_dir.mkdir(parents=True)
-        script_file = scripts_dir / "deepseek_v3.sh"
+        script_file = scripts_dir / "deepseek_v3.py"
         script_file.touch()
 
         monkeypatch.setattr(
diff --git a/tests/fast/utils/debug_utils/run_megatron/cli/test_worker_executor.py b/tests/fast/utils/debug_utils/run_megatron/cli/test_worker_executor.py
index ea0aa24d8d6..27f4c3929aa 100644
--- a/tests/fast/utils/debug_utils/run_megatron/cli/test_worker_executor.py
+++ b/tests/fast/utils/debug_utils/run_megatron/cli/test_worker_executor.py
@@ -182,9 +182,9 @@ def test_no_routing_replay(self) -> None:
 
 
 class TestBuildTorchrunCmd:
-    @patch("miles.utils.debug_utils.run_megatron.cli.worker_executor.resolve_model_script")
-    def test_basic_structure(self, mock_resolve: object) -> None:
-        mock_resolve.return_value = Path("/repo/scripts/models/deepseek_v3.sh")  # type: ignore[union-attr]
+    @patch("miles.utils.debug_utils.run_megatron.cli.worker_executor.load_model_args")
+    def test_basic_structure(self, mock_load: object) -> None:
+        mock_load.return_value = "--num-layers 61"  # type: ignore[union-attr]
         cmd = build_torchrun_cmd(
             model_type="deepseek_v3",
             megatron_path=Path("/megatron"),
@@ -192,12 +192,12 @@ def test_basic_structure(self, mock_resolve: object) -> None:
             worker_args="--foo bar",
         )
         assert "torchrun" in cmd
-        assert "source" in cmd
+        assert "--num-layers 61" in cmd
         assert "PYTHONPATH" in cmd
 
-    @patch("miles.utils.debug_utils.run_megatron.cli.worker_executor.resolve_model_script")
-    def test_nproc(self, mock_resolve: object) -> None:
-        mock_resolve.return_value = Path("/repo/scripts/models/test.sh")  # type: ignore[union-attr]
+    @patch("miles.utils.debug_utils.run_megatron.cli.worker_executor.load_model_args")
+    def test_nproc(self, mock_load: object) -> None:
+        mock_load.return_value = "--num-layers 61"  # type: ignore[union-attr]
         cmd = build_torchrun_cmd(
             model_type="test",
             megatron_path=Path("/megatron"),
@@ -206,9 +206,9 @@ def test_nproc(self, mock_resolve: object) -> None:
         )
         assert "--nproc-per-node 8" in cmd
 
-    @patch("miles.utils.debug_utils.run_megatron.cli.worker_executor.resolve_model_script")
-    def test_worker_args_in_cmd(self, mock_resolve: object) -> None:
-        mock_resolve.return_value = Path("/repo/scripts/models/test.sh")  # type: ignore[union-attr]
+    @patch("miles.utils.debug_utils.run_megatron.cli.worker_executor.load_model_args")
+    def test_worker_args_in_cmd(self, mock_load: object) -> None:
+        mock_load.return_value = "--num-layers 61"  # type: ignore[union-attr]
         cmd = build_torchrun_cmd(
             model_type="test",
             megatron_path=Path("/megatron"),
@@ -217,9 +217,9 @@ def test_worker_args_in_cmd(self, mock_resolve: object) -> None:
         )
         assert "--my-flag 42" in cmd
 
-    @patch("miles.utils.debug_utils.run_megatron.cli.worker_executor.resolve_model_script")
-    def test_megatron_in_pythonpath(self, mock_resolve: object) -> None:
-        mock_resolve.return_value = Path("/repo/scripts/models/test.sh")  # type: ignore[union-attr]
+    @patch("miles.utils.debug_utils.run_megatron.cli.worker_executor.load_model_args")
+    def test_megatron_in_pythonpath(self, mock_load: object) -> None:
+        mock_load.return_value = "--num-layers 61"  # type: ignore[union-attr]
         cmd = build_torchrun_cmd(
             model_type="test",
             megatron_path=Path("/my/megatron"),
diff --git a/tests/fast/utils/external_utils/test_model_args_utils.py b/tests/fast/utils/external_utils/test_model_args_utils.py
new file mode 100644
index 00000000000..3b6fece1c6f
--- /dev/null
+++ b/tests/fast/utils/external_utils/test_model_args_utils.py
@@ -0,0 +1,233 @@
+import os
+import subprocess
+import sys
+from pathlib import Path
+
+import pytest
+
+import miles.utils.external_utils.model_args_utils as model_args_utils
+from miles.utils.external_utils.model_args_utils import load_model_args, load_sibling_model_args, moe_layer_freq
+
+_MODEL_ARGS_CLI = Path(model_args_utils.__file__).resolve()
+
+_SCRIPT_BODY = """
+import os
+
+from model_args_utils import moe_layer_freq
+
+
+def model_args(nlayers: int | None = None) -> str:
+    nlayers = nlayers if nlayers is not None else int(os.environ.get("MODEL_ARGS_NUM_LAYERS") or 61)
+    return (
+        "--swiglu "
+        f"--num-layers {nlayers} "
+        f"--moe-layer-freq {moe_layer_freq(nlayers=nlayers, first_k_dense_replace=3)} "
+    )
+"""
+
+_WRAPPER_BODY = """
+from model_args_utils import load_sibling_model_args
+
+
+def model_args() -> str:
+    return load_sibling_model_args(__file__, "fake-model.4layer", nlayers=4)
+"""
+
+
+@pytest.fixture
+def model_script(monkeypatch, tmp_path):
+    path = tmp_path / "fake-model.4layer.py"
+    path.write_text(_SCRIPT_BODY)
+    (tmp_path / "fake-wrapper.py").write_text(_WRAPPER_BODY)
+    monkeypatch.setattr("miles.utils.external_utils.model_args_utils.MODEL_SCRIPT_DIR", tmp_path)
+    monkeypatch.delenv("MODEL_ARGS_NUM_LAYERS", raising=False)
+    return path
+
+
+class TestLoadModelArgsSplitting:
+    def test_splits_each_line_the_way_read_ra_would(self, model_script):
+        """One source line per flag must expand to the same argv the shell array held."""
+        assert load_model_args("fake-model.4layer").split()[:3] == ["--swiglu", "--num-layers", "61"]
+
+    def test_collapses_a_declaration_that_spans_several_lines(self, model_script):
+        """A newline makes the launcher's `read -ra ... <<< "$(...)"` stop after the first line, silently."""
+        model_script.write_text("def model_args() -> str:\n    return '--a 1\\n--b 2'\n")
+
+        assert load_model_args("fake-model.4layer") == "--a 1 --b 2"
+
+    def test_rejects_a_model_script_that_declares_nothing(self, model_script):
+        """An all-whitespace declaration is a generator bug, not an empty argument."""
+        model_script.write_text("def model_args() -> str:\n    return '   '\n")
+
+        with pytest.raises(AssertionError):
+            load_model_args("fake-model.4layer")
+
+    def test_keeps_the_bracket_patterns_megatron_expects(self, model_script):
+        """--moe-layer-freq values contain brackets and stars, which must survive as one token."""
+        model_script.write_text("def model_args() -> str:\n    return '--moe-layer-freq [0]*3+[1]*75'\n")
+
+        assert load_model_args("fake-model.4layer") == "--moe-layer-freq [0]*3+[1]*75"
+
+
+class TestMoeLayerFreq:
+    def test_renders_the_dense_prefix_then_moe_layers(self):
+        """The mask must match `arr+=(0)` for the first K layers and `arr+=(1)` after."""
+        assert moe_layer_freq(nlayers=5, first_k_dense_replace=2) == "[0,0,1,1,1]"
+
+    def test_renders_only_dense_layers_when_the_model_is_shorter_than_the_dense_prefix(self):
+        """The shell loop ran over the layer count, so a 2-layer deepseek-v3 got [0,0], not [0,0,0]."""
+        assert moe_layer_freq(nlayers=2, first_k_dense_replace=3) == "[0,0]"
+
+    def test_renders_an_all_moe_mask_when_no_dense_layers(self):
+        """DeepSeek V4 has no dense prefix, so every entry is a MoE layer."""
+        assert moe_layer_freq(nlayers=3, first_k_dense_replace=0) == "[1,1,1]"
+
+
+class TestLoadModelArgs:
+    def test_returns_the_declared_argv(self, model_script):
+        """A python consumer gets argv tokens directly instead of sourcing a shell script."""
+        assert load_model_args("fake-model.4layer").split() == [
+            "--swiglu",
+            "--num-layers",
+            "61",
+            "--moe-layer-freq",
+            moe_layer_freq(nlayers=61, first_k_dense_replace=3),
+        ]
+
+    def test_forwards_keyword_overrides(self, model_script):
+        """Layer-count variants are the same script called with a different argument."""
+        assert load_model_args("fake-model.4layer", nlayers=4).split()[2] == "4"
+
+    def test_rejects_an_unknown_model_type(self, model_script):
+        """A typo must fail loudly rather than silently produce an argument-less run."""
+        with pytest.raises(AssertionError):
+            load_model_args("no-such-model")
+
+    def test_a_dotted_filename_is_importable(self, model_script):
+        """Model names like glm4.5-106B-A12B cannot be imported by module path."""
+        assert "." in model_script.stem
+        assert load_model_args(model_script.stem)
+
+    def test_reads_the_model_scripts_of_the_requested_checkout(self, model_script, tmp_path_factory):
+        """A launcher must get the model definition of its own checkout, not of the installed package."""
+        other = tmp_path_factory.mktemp("other-checkout")
+        (other / model_script.name).write_text(
+            _SCRIPT_BODY.replace('"MODEL_ARGS_NUM_LAYERS") or 61', '"UNUSED") or 7')
+        )
+
+        assert load_model_args(model_script.stem, model_script_dir=other).split()[2] == "7"
+
+    def test_a_wrapper_stays_inside_the_checkout_it_was_loaded_from(self, model_script, tmp_path_factory):
+        """A variant script must reach the base script next to it, not the one of the installed package."""
+        other = tmp_path_factory.mktemp("other-checkout")
+        (other / "fake-wrapper.py").write_text(_WRAPPER_BODY)
+        (other / model_script.name).write_text(
+            _SCRIPT_BODY.replace('"MODEL_ARGS_NUM_LAYERS") or 61', '"UNUSED") or 7')
+        )
+
+        assert load_model_args("fake-wrapper", model_script_dir=other).split() == ["--swiglu", "--num-layers", "4"] + [
+            "--moe-layer-freq",
+            moe_layer_freq(nlayers=4, first_k_dense_replace=3),
+        ]
+
+    def test_still_honours_the_environment_override_the_shell_scripts_read(self, model_script, monkeypatch):
+        """MODEL_ARGS_NUM_LAYERS used to reach the sourced .sh, so it must reach the .py too."""
+        monkeypatch.setenv("MODEL_ARGS_NUM_LAYERS", "9")
+
+        assert load_model_args("fake-model.4layer").split()[2] == "9"
+
+    def test_ignores_an_empty_environment_override(self, model_script, monkeypatch):
+        """`${VAR:-default}` falls back to the default when the variable is set but empty."""
+        monkeypatch.setenv("MODEL_ARGS_NUM_LAYERS", "")
+
+        assert load_model_args("fake-model.4layer").split()[2] == "61"
+
+    def test_an_explicit_override_beats_the_environment(self, model_script, monkeypatch):
+        """`MODEL_ARGS_NUM_LAYERS=5 source x.sh` let the caller's assignment win; keyword arguments must too."""
+        monkeypatch.setenv("MODEL_ARGS_NUM_LAYERS", "9")
+
+        assert load_model_args("fake-model.4layer", nlayers=4).split()[2] == "4"
+
+    def test_honours_a_zero_override(self, model_script):
+        """A dense-layer count of zero is a real value; `x or default` would silently restore the default."""
+        model_script.write_text(_SCRIPT_BODY.replace("first_k_dense_replace=3", "first_k_dense_replace=nlayers"))
+
+        assert load_model_args("fake-model.4layer", nlayers=0) == "--swiglu --num-layers 0 --moe-layer-freq []"
+
+    def test_ignores_an_environment_override_the_model_does_not_declare(self, model_script, monkeypatch):
+        """A model without a rotary base must not fail because some other model's variable is exported."""
+        monkeypatch.setenv("MODEL_ARGS_ROTARY_BASE", "5000000")
+
+        assert load_model_args("fake-model.4layer").split()[2] == "61"
+
+    def test_rejects_an_override_the_model_does_not_declare(self, model_script):
+        """The keyword reaches model_args() directly, so a misspelling is a TypeError rather than a silent no-op."""
+        with pytest.raises(TypeError):
+            load_model_args("fake-model.4layer", n_layers=4)
+
+
+class TestLoadSiblingModelArgs:
+    def test_resolves_the_base_next_to_the_variant_script(self, model_script, tmp_path_factory):
+        """The variant knows where it lives; nothing else in the process does."""
+        other = tmp_path_factory.mktemp("sibling-checkout")
+        (other / model_script.name).write_text(
+            _SCRIPT_BODY.replace('"MODEL_ARGS_NUM_LAYERS") or 61', '"UNUSED") or 7')
+        )
+
+        assert load_sibling_model_args(str(other / "anything.py"), model_script.stem).split()[2] == "7"
+
+
+class TestModelArgsScript:
+    def test_shell_consumers_recover_the_original_tokens(self):
+        """Bracket patterns must survive read -ra without being glob-expanded."""
+        script = (
+            f'set -e; MODEL_ARGS_LINE="$({sys.executable} {_MODEL_ARGS_CLI} qwen3-4B)" || exit 1; '
+            'read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"; printf "%s\\n" "${MODEL_ARGS[@]}"'
+        )
+        result = subprocess.run(["bash", "-c", script], capture_output=True, text=True, check=True)
+
+        assert result.stdout.splitlines() == load_model_args("qwen3-4B").split()
+
+    def test_an_unknown_model_type_stops_the_launcher(self):
+        """A bare `read -ra ... <<< "$(...)"` swallows the failure and trains with no architecture flags."""
+        script = (
+            f'MODEL_ARGS_LINE="$({sys.executable} {_MODEL_ARGS_CLI} no-such-model 2>/dev/null)" || exit 1; '
+            'echo "the launcher kept going"'
+        )
+        result = subprocess.run(["bash", "-c", script], capture_output=True, text=True)
+
+        assert result.returncode == 1
+        assert result.stdout == ""
+
+    def test_a_here_string_read_stops_at_the_first_line(self):
+        """Why load_model_args() collapses its result: read -ra drops the rest of a multi-line value silently."""
+        script = 'read -ra MODEL_ARGS <<< "$1"; printf "%s\\n" "${MODEL_ARGS[@]}"'
+        result = subprocess.run(
+            ["bash", "-c", script, "_", "--a 1\n--b 2"], capture_output=True, text=True, check=True
+        )
+
+        assert result.stdout.split() == ["--a", "1"]
+
+    def test_runs_from_a_checkout_whose_package_is_not_installed(self):
+        """Executed by path with no site-packages, a model script must still reach the loader's helpers."""
+        result = subprocess.run(
+            [sys.executable, "-S", "-E", str(_MODEL_ARGS_CLI), "qwen3-30B-A3B"],
+            capture_output=True,
+            text=True,
+            check=True,
+            cwd="/",
+        )
+
+        assert "--num-layers" in result.stdout
+
+    def test_forwards_the_rotary_base_override(self):
+        """The geo3k launcher prefixes the command with MODEL_ARGS_ROTARY_BASE, as the shell scripts did."""
+        result = subprocess.run(
+            [sys.executable, str(_MODEL_ARGS_CLI), "qwen3-4B"],
+            capture_output=True,
+            text=True,
+            check=True,
+            env={**os.environ, "MODEL_ARGS_ROTARY_BASE": "5000000"},
+        )
+
+        assert "--rotary-base 5000000" in result.stdout
diff --git a/tests/snapshots/launch_scripts/self_executing/examples/experimental/formal_math/single_round/run_minimal.py/import.txt b/tests/snapshots/launch_scripts/self_executing/examples/experimental/formal_math/single_round/run_minimal.py/import.txt
index ad94fb21e00..04337367e57 100644
--- a/tests/snapshots/launch_scripts/self_executing/examples/experimental/formal_math/single_round/run_minimal.py/import.txt
+++ b/tests/snapshots/launch_scripts/self_executing/examples/experimental/formal_math/single_round/run_minimal.py/import.txt
@@ -1,8 +1,24 @@
 ### 0
-bash -c export PYTHONUNBUFFERED=1 && source "<REPO_ROOT>/scripts/models/qwen3-8B.sh" && ray job submit
+bash -c export PYTHONUNBUFFERED=1 && ray job submit
   --address="http://127.0.0.1:8265"
   --runtime-env-json='{"env_vars": {"PYTHONUNBUFFERED": "1", "PYTHONPATH": "/root/Megatron-LM/", "CUDA_DEVICE_MAX_CONNECTIONS": "1", "NCCL_NVLS_ENABLE": "1"}}'
-  -- python3 train.py ${MODEL_ARGS[@]}
+  -- python3 train.py
+  --swiglu
+  --num-layers 36
+  --hidden-size 4096
+  --ffn-hidden-size 12288
+  --num-attention-heads 32
+  --group-query-attention
+  --num-query-groups 8
+  --use-rotary-position-embeddings
+  --disable-bias-linear
+  --normalization RMSNorm
+  --norm-epsilon 1e-6
+  --rotary-base 1000000
+  --vocab-size 151936
+  --kv-channels 128
+  --qk-layernorm
+  --untie-embeddings-and-output-weights
   --hf-checkpoint /root/models/Qwen3-8B/
   --ref-load /root/models/Qwen3-8B_torch_dist
   --save-interval 20
diff --git a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-4.5-Air/broadcast.txt b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-4.5-Air/broadcast.txt
index 4a24915d1fb..316ebbc2213 100644
--- a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-4.5-Air/broadcast.txt
+++ b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-4.5-Air/broadcast.txt
@@ -39,7 +39,7 @@ RAY_memory_monitor_refresh_ms=0 ray start
 python3 -c "import ray; ray.init(address='auto', ignore_reinit_error=True); print(int(ray.cluster_resources().get('GPU', 0))); ray.shutdown()"
 
 ### 11
-MODEL_ARGS_ROTARY_BASE=1000000 source "<REPO_ROOT>/scripts/models/glm4.5-106B-A12B.sh" && ray job submit
+MODEL_ARGS_LINE="$(MODEL_ARGS_ROTARY_BASE=1000000 python3 "<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py" glm4.5-106B-A12B)" || exit 1; read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}" && ray job submit
   --address='http://127.0.0.1:8265'
   --runtime-env-json='{"env_vars": {"PYTHONUNBUFFERED": "1", "RAY_DEBUG": "1", "PYTHONPATH": "/root/Megatron-LM/", "CUDA_DEVICE_MAX_CONNECTIONS": "1", "NCCL_NVLS_ENABLE": "0", "MILES_LOG_DIR": "<SANDBOX>", "MC_TRANSFER_TIMEOUT": "300", "MODEL_ARGS_ROTARY_BASE": "1000000"}}'
   -- python3 "<REPO_ROOT>/train.py" ${MODEL_ARGS[@]}
diff --git a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-4.5-Air/p2p.txt b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-4.5-Air/p2p.txt
index b992f747747..a839b640b95 100644
--- a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-4.5-Air/p2p.txt
+++ b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-4.5-Air/p2p.txt
@@ -39,7 +39,7 @@ RAY_memory_monitor_refresh_ms=0 ray start
 python3 -c "import ray; ray.init(address='auto', ignore_reinit_error=True); print(int(ray.cluster_resources().get('GPU', 0))); ray.shutdown()"
 
 ### 11
-MODEL_ARGS_ROTARY_BASE=1000000 source "<REPO_ROOT>/scripts/models/glm4.5-106B-A12B.sh" && ray job submit
+MODEL_ARGS_LINE="$(MODEL_ARGS_ROTARY_BASE=1000000 python3 "<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py" glm4.5-106B-A12B)" || exit 1; read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}" && ray job submit
   --address='http://127.0.0.1:8265'
   --runtime-env-json='{"env_vars": {"PYTHONUNBUFFERED": "1", "RAY_DEBUG": "1", "PYTHONPATH": "/root/Megatron-LM/", "CUDA_DEVICE_MAX_CONNECTIONS": "1", "NCCL_NVLS_ENABLE": "0", "MILES_LOG_DIR": "<SANDBOX>", "MC_TRANSFER_TIMEOUT": "300", "MODEL_ARGS_ROTARY_BASE": "1000000"}}'
   -- python3 "<REPO_ROOT>/train.py" ${MODEL_ARGS[@]}
diff --git a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-4.7-Flash/broadcast.txt b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-4.7-Flash/broadcast.txt
index 433f05bf261..c4dec6d382e 100644
--- a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-4.7-Flash/broadcast.txt
+++ b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-4.7-Flash/broadcast.txt
@@ -39,7 +39,7 @@ RAY_memory_monitor_refresh_ms=0 ray start
 python3 -c "import ray; ray.init(address='auto', ignore_reinit_error=True); print(int(ray.cluster_resources().get('GPU', 0))); ray.shutdown()"
 
 ### 11
-source "<REPO_ROOT>/scripts/models/glm4.7-flash.sh" && ray job submit
+MODEL_ARGS_LINE="$(python3 "<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py" glm4.7-flash)" || exit 1; read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}" && ray job submit
   --address='http://127.0.0.1:8265'
   --runtime-env-json='{"env_vars": {"PYTHONUNBUFFERED": "1", "RAY_DEBUG": "1", "PYTHONPATH": "/root/Megatron-LM/", "CUDA_DEVICE_MAX_CONNECTIONS": "1", "NCCL_NVLS_ENABLE": "0", "MILES_LOG_DIR": "<SANDBOX>", "MC_TRANSFER_TIMEOUT": "300"}}'
   -- python3 "<REPO_ROOT>/train.py" ${MODEL_ARGS[@]}
diff --git a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-4.7-Flash/p2p.txt b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-4.7-Flash/p2p.txt
index fd2c92f05f4..0758e1a264e 100644
--- a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-4.7-Flash/p2p.txt
+++ b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-4.7-Flash/p2p.txt
@@ -39,7 +39,7 @@ RAY_memory_monitor_refresh_ms=0 ray start
 python3 -c "import ray; ray.init(address='auto', ignore_reinit_error=True); print(int(ray.cluster_resources().get('GPU', 0))); ray.shutdown()"
 
 ### 11
-source "<REPO_ROOT>/scripts/models/glm4.7-flash.sh" && ray job submit
+MODEL_ARGS_LINE="$(python3 "<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py" glm4.7-flash)" || exit 1; read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}" && ray job submit
   --address='http://127.0.0.1:8265'
   --runtime-env-json='{"env_vars": {"PYTHONUNBUFFERED": "1", "RAY_DEBUG": "1", "PYTHONPATH": "/root/Megatron-LM/", "CUDA_DEVICE_MAX_CONNECTIONS": "1", "NCCL_NVLS_ENABLE": "0", "MILES_LOG_DIR": "<SANDBOX>", "MC_TRANSFER_TIMEOUT": "300"}}'
   -- python3 "<REPO_ROOT>/train.py" ${MODEL_ARGS[@]}
diff --git a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-5/broadcast.txt b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-5/broadcast.txt
index 878aae279ca..6a1821a5298 100644
--- a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-5/broadcast.txt
+++ b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-5/broadcast.txt
@@ -39,7 +39,7 @@ RAY_memory_monitor_refresh_ms=0 ray start
 python3 -c "import ray; ray.init(address='auto', ignore_reinit_error=True); print(int(ray.cluster_resources().get('GPU', 0))); ray.shutdown()"
 
 ### 11
-source "<REPO_ROOT>/scripts/models/glm5-744B-A40B.sh" && ray job submit
+MODEL_ARGS_LINE="$(python3 "<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py" glm5-744B-A40B)" || exit 1; read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}" && ray job submit
   --address='http://127.0.0.1:8265'
   --runtime-env-json='{"env_vars": {"PYTHONUNBUFFERED": "1", "RAY_DEBUG": "1", "PYTHONPATH": "/root/Megatron-LM/", "CUDA_DEVICE_MAX_CONNECTIONS": "1", "NCCL_NVLS_ENABLE": "1", "MILES_LOG_DIR": "<SANDBOX>", "MC_TRANSFER_TIMEOUT": "600", "INDEXER_ROPE_NEOX_STYLE": "0", "NVSHMEM_DISABLE_NCCL": "1"}}'
   -- python3 "<REPO_ROOT>/train.py" ${MODEL_ARGS[@]}
diff --git a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-5/p2p.txt b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-5/p2p.txt
index 84ba31d3922..8a9ec08070c 100644
--- a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-5/p2p.txt
+++ b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-5/p2p.txt
@@ -39,7 +39,7 @@ RAY_memory_monitor_refresh_ms=0 ray start
 python3 -c "import ray; ray.init(address='auto', ignore_reinit_error=True); print(int(ray.cluster_resources().get('GPU', 0))); ray.shutdown()"
 
 ### 11
-source "<REPO_ROOT>/scripts/models/glm5-744B-A40B.sh" && ray job submit
+MODEL_ARGS_LINE="$(python3 "<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py" glm5-744B-A40B)" || exit 1; read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}" && ray job submit
   --address='http://127.0.0.1:8265'
   --runtime-env-json='{"env_vars": {"PYTHONUNBUFFERED": "1", "RAY_DEBUG": "1", "PYTHONPATH": "/root/Megatron-LM/", "CUDA_DEVICE_MAX_CONNECTIONS": "1", "NCCL_NVLS_ENABLE": "1", "MILES_LOG_DIR": "<SANDBOX>", "MC_TRANSFER_TIMEOUT": "600", "INDEXER_ROPE_NEOX_STYLE": "0", "NVSHMEM_DISABLE_NCCL": "1"}}'
   -- python3 "<REPO_ROOT>/train.py" ${MODEL_ARGS[@]}
diff --git a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-5_20layer/broadcast.txt b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-5_20layer/broadcast.txt
index f2e0345b5d0..66c44a8f745 100644
--- a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-5_20layer/broadcast.txt
+++ b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-5_20layer/broadcast.txt
@@ -39,7 +39,7 @@ RAY_memory_monitor_refresh_ms=0 ray start
 python3 -c "import ray; ray.init(address='auto', ignore_reinit_error=True); print(int(ray.cluster_resources().get('GPU', 0))); ray.shutdown()"
 
 ### 11
-source "<REPO_ROOT>/scripts/models/glm5-744B-A40B_20layer.sh" && ray job submit
+MODEL_ARGS_LINE="$(python3 "<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py" glm5-744B-A40B_20layer)" || exit 1; read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}" && ray job submit
   --address='http://127.0.0.1:8265'
   --runtime-env-json='{"env_vars": {"PYTHONUNBUFFERED": "1", "RAY_DEBUG": "1", "PYTHONPATH": "/root/Megatron-LM/", "CUDA_DEVICE_MAX_CONNECTIONS": "1", "NCCL_NVLS_ENABLE": "1", "MILES_LOG_DIR": "<SANDBOX>", "MC_TRANSFER_TIMEOUT": "600", "INDEXER_ROPE_NEOX_STYLE": "0", "NVSHMEM_DISABLE_NCCL": "1"}}'
   -- python3 "<REPO_ROOT>/train.py" ${MODEL_ARGS[@]}
diff --git a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-5_20layer/p2p.txt b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-5_20layer/p2p.txt
index 060b1c407c5..8985ffe0ed3 100644
--- a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-5_20layer/p2p.txt
+++ b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-5_20layer/p2p.txt
@@ -39,7 +39,7 @@ RAY_memory_monitor_refresh_ms=0 ray start
 python3 -c "import ray; ray.init(address='auto', ignore_reinit_error=True); print(int(ray.cluster_resources().get('GPU', 0))); ray.shutdown()"
 
 ### 11
-source "<REPO_ROOT>/scripts/models/glm5-744B-A40B_20layer.sh" && ray job submit
+MODEL_ARGS_LINE="$(python3 "<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py" glm5-744B-A40B_20layer)" || exit 1; read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}" && ray job submit
   --address='http://127.0.0.1:8265'
   --runtime-env-json='{"env_vars": {"PYTHONUNBUFFERED": "1", "RAY_DEBUG": "1", "PYTHONPATH": "/root/Megatron-LM/", "CUDA_DEVICE_MAX_CONNECTIONS": "1", "NCCL_NVLS_ENABLE": "1", "MILES_LOG_DIR": "<SANDBOX>", "MC_TRANSFER_TIMEOUT": "600", "INDEXER_ROPE_NEOX_STYLE": "0", "NVSHMEM_DISABLE_NCCL": "1"}}'
   -- python3 "<REPO_ROOT>/train.py" ${MODEL_ARGS[@]}
diff --git a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-5_4layer/broadcast.txt b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-5_4layer/broadcast.txt
index 76fdfe325cd..1ee2b15c8e1 100644
--- a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-5_4layer/broadcast.txt
+++ b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-5_4layer/broadcast.txt
@@ -39,7 +39,7 @@ RAY_memory_monitor_refresh_ms=0 ray start
 python3 -c "import ray; ray.init(address='auto', ignore_reinit_error=True); print(int(ray.cluster_resources().get('GPU', 0))); ray.shutdown()"
 
 ### 11
-source "<REPO_ROOT>/scripts/models/glm5-744B-A40B_4layer.sh" && ray job submit
+MODEL_ARGS_LINE="$(python3 "<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py" glm5-744B-A40B_4layer)" || exit 1; read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}" && ray job submit
   --address='http://127.0.0.1:8265'
   --runtime-env-json='{"env_vars": {"PYTHONUNBUFFERED": "1", "RAY_DEBUG": "1", "PYTHONPATH": "/root/Megatron-LM/", "CUDA_DEVICE_MAX_CONNECTIONS": "1", "NCCL_NVLS_ENABLE": "1", "MILES_LOG_DIR": "<SANDBOX>", "MC_TRANSFER_TIMEOUT": "600", "INDEXER_ROPE_NEOX_STYLE": "0", "NVSHMEM_DISABLE_NCCL": "1"}}'
   -- python3 "<REPO_ROOT>/train.py" ${MODEL_ARGS[@]}
diff --git a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-5_4layer/p2p.txt b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-5_4layer/p2p.txt
index d1012211f08..51b11f3c7c5 100644
--- a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-5_4layer/p2p.txt
+++ b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-5_4layer/p2p.txt
@@ -39,7 +39,7 @@ RAY_memory_monitor_refresh_ms=0 ray start
 python3 -c "import ray; ray.init(address='auto', ignore_reinit_error=True); print(int(ray.cluster_resources().get('GPU', 0))); ray.shutdown()"
 
 ### 11
-source "<REPO_ROOT>/scripts/models/glm5-744B-A40B_4layer.sh" && ray job submit
+MODEL_ARGS_LINE="$(python3 "<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py" glm5-744B-A40B_4layer)" || exit 1; read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}" && ray job submit
   --address='http://127.0.0.1:8265'
   --runtime-env-json='{"env_vars": {"PYTHONUNBUFFERED": "1", "RAY_DEBUG": "1", "PYTHONPATH": "/root/Megatron-LM/", "CUDA_DEVICE_MAX_CONNECTIONS": "1", "NCCL_NVLS_ENABLE": "1", "MILES_LOG_DIR": "<SANDBOX>", "MC_TRANSFER_TIMEOUT": "600", "INDEXER_ROPE_NEOX_STYLE": "0", "NVSHMEM_DISABLE_NCCL": "1"}}'
   -- python3 "<REPO_ROOT>/train.py" ${MODEL_ARGS[@]}
diff --git a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-Z1-9B-0414/broadcast.txt b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-Z1-9B-0414/broadcast.txt
index 560206d4205..98d9bec5996 100644
--- a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-Z1-9B-0414/broadcast.txt
+++ b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-Z1-9B-0414/broadcast.txt
@@ -35,7 +35,7 @@ ray start
   --dashboard-port=8265
 
 ### 10
-source "<REPO_ROOT>/scripts/models/glm4-9B.sh" && ray job submit
+MODEL_ARGS_LINE="$(python3 "<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py" glm4-9B)" || exit 1; read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}" && ray job submit
   --address='http://127.0.0.1:8265'
   --runtime-env-json='{"env_vars": {"PYTHONUNBUFFERED": "1", "RAY_DEBUG": "1", "PYTHONPATH": "/root/Megatron-LM/", "CUDA_DEVICE_MAX_CONNECTIONS": "1", "NCCL_NVLS_ENABLE": "0", "MILES_LOG_DIR": "<SANDBOX>"}}'
   -- python3 "<REPO_ROOT>/train.py" ${MODEL_ARGS[@]}
diff --git a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-Z1-9B-0414/p2p.txt b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-Z1-9B-0414/p2p.txt
index cb2a9a61cbc..f05548c663d 100644
--- a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-Z1-9B-0414/p2p.txt
+++ b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/GLM-Z1-9B-0414/p2p.txt
@@ -35,7 +35,7 @@ ray start
   --dashboard-port=8265
 
 ### 10
-source "<REPO_ROOT>/scripts/models/glm4-9B.sh" && ray job submit
+MODEL_ARGS_LINE="$(python3 "<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py" glm4-9B)" || exit 1; read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}" && ray job submit
   --address='http://127.0.0.1:8265'
   --runtime-env-json='{"env_vars": {"PYTHONUNBUFFERED": "1", "RAY_DEBUG": "1", "PYTHONPATH": "/root/Megatron-LM/", "CUDA_DEVICE_MAX_CONNECTIONS": "1", "NCCL_NVLS_ENABLE": "0", "MILES_LOG_DIR": "<SANDBOX>"}}'
   -- python3 "<REPO_ROOT>/train.py" ${MODEL_ARGS[@]}
diff --git a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Kimi-K2-Instruct/broadcast.txt b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Kimi-K2-Instruct/broadcast.txt
index 115505f9e0f..335c01146d1 100644
--- a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Kimi-K2-Instruct/broadcast.txt
+++ b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Kimi-K2-Instruct/broadcast.txt
@@ -39,7 +39,7 @@ RAY_memory_monitor_refresh_ms=0 ray start
 python3 -c "import ray; ray.init(address='auto', ignore_reinit_error=True); print(int(ray.cluster_resources().get('GPU', 0))); ray.shutdown()"
 
 ### 11
-source "<REPO_ROOT>/scripts/models/kimi-k2.sh" && ray job submit
+MODEL_ARGS_LINE="$(python3 "<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py" kimi-k2)" || exit 1; read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}" && ray job submit
   --address='http://127.0.0.1:8265'
   --runtime-env-json='{"env_vars": {"PYTHONUNBUFFERED": "1", "RAY_DEBUG": "1", "PYTHONPATH": "/root/Megatron-LM/", "CUDA_DEVICE_MAX_CONNECTIONS": "1", "NCCL_NVLS_ENABLE": "1", "MILES_LOG_DIR": "<SANDBOX>", "MC_TRANSFER_TIMEOUT": "300"}}'
   -- python3 "<REPO_ROOT>/train.py" ${MODEL_ARGS[@]}
diff --git a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Kimi-K2-Instruct/p2p.txt b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Kimi-K2-Instruct/p2p.txt
index 52de614a9da..31061377330 100644
--- a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Kimi-K2-Instruct/p2p.txt
+++ b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Kimi-K2-Instruct/p2p.txt
@@ -39,7 +39,7 @@ RAY_memory_monitor_refresh_ms=0 ray start
 python3 -c "import ray; ray.init(address='auto', ignore_reinit_error=True); print(int(ray.cluster_resources().get('GPU', 0))); ray.shutdown()"
 
 ### 11
-source "<REPO_ROOT>/scripts/models/kimi-k2.sh" && ray job submit
+MODEL_ARGS_LINE="$(python3 "<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py" kimi-k2)" || exit 1; read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}" && ray job submit
   --address='http://127.0.0.1:8265'
   --runtime-env-json='{"env_vars": {"PYTHONUNBUFFERED": "1", "RAY_DEBUG": "1", "PYTHONPATH": "/root/Megatron-LM/", "CUDA_DEVICE_MAX_CONNECTIONS": "1", "NCCL_NVLS_ENABLE": "1", "MILES_LOG_DIR": "<SANDBOX>", "MC_TRANSFER_TIMEOUT": "300"}}'
   -- python3 "<REPO_ROOT>/train.py" ${MODEL_ARGS[@]}
diff --git a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Moonlight-16B-A3B-Instruct/broadcast.txt b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Moonlight-16B-A3B-Instruct/broadcast.txt
index ee975b6212c..4b5032c1c73 100644
--- a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Moonlight-16B-A3B-Instruct/broadcast.txt
+++ b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Moonlight-16B-A3B-Instruct/broadcast.txt
@@ -39,7 +39,7 @@ RAY_memory_monitor_refresh_ms=0 ray start
 python3 -c "import ray; ray.init(address='auto', ignore_reinit_error=True); print(int(ray.cluster_resources().get('GPU', 0))); ray.shutdown()"
 
 ### 11
-source "<REPO_ROOT>/scripts/models/moonlight.sh" && ray job submit
+MODEL_ARGS_LINE="$(python3 "<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py" moonlight)" || exit 1; read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}" && ray job submit
   --address='http://127.0.0.1:8265'
   --runtime-env-json='{"env_vars": {"PYTHONUNBUFFERED": "1", "RAY_DEBUG": "1", "PYTHONPATH": "/root/Megatron-LM/", "CUDA_DEVICE_MAX_CONNECTIONS": "1", "NCCL_NVLS_ENABLE": "0", "MILES_LOG_DIR": "<SANDBOX>", "MC_TRANSFER_TIMEOUT": "300"}}'
   -- python3 "<REPO_ROOT>/train.py" ${MODEL_ARGS[@]}
diff --git a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Moonlight-16B-A3B-Instruct/p2p.txt b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Moonlight-16B-A3B-Instruct/p2p.txt
index 5886226820a..5fa21e2e528 100644
--- a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Moonlight-16B-A3B-Instruct/p2p.txt
+++ b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Moonlight-16B-A3B-Instruct/p2p.txt
@@ -39,7 +39,7 @@ RAY_memory_monitor_refresh_ms=0 ray start
 python3 -c "import ray; ray.init(address='auto', ignore_reinit_error=True); print(int(ray.cluster_resources().get('GPU', 0))); ray.shutdown()"
 
 ### 11
-source "<REPO_ROOT>/scripts/models/moonlight.sh" && ray job submit
+MODEL_ARGS_LINE="$(python3 "<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py" moonlight)" || exit 1; read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}" && ray job submit
   --address='http://127.0.0.1:8265'
   --runtime-env-json='{"env_vars": {"PYTHONUNBUFFERED": "1", "RAY_DEBUG": "1", "PYTHONPATH": "/root/Megatron-LM/", "CUDA_DEVICE_MAX_CONNECTIONS": "1", "NCCL_NVLS_ENABLE": "0", "MILES_LOG_DIR": "<SANDBOX>", "MC_TRANSFER_TIMEOUT": "300"}}'
   -- python3 "<REPO_ROOT>/train.py" ${MODEL_ARGS[@]}
diff --git a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Qwen3-235B-A22B-Instruct-2507/broadcast.txt b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Qwen3-235B-A22B-Instruct-2507/broadcast.txt
index 225afb2058d..19721090d74 100644
--- a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Qwen3-235B-A22B-Instruct-2507/broadcast.txt
+++ b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Qwen3-235B-A22B-Instruct-2507/broadcast.txt
@@ -39,7 +39,7 @@ RAY_memory_monitor_refresh_ms=0 ray start
 python3 -c "import ray; ray.init(address='auto', ignore_reinit_error=True); print(int(ray.cluster_resources().get('GPU', 0))); ray.shutdown()"
 
 ### 11
-MODEL_ARGS_ROTARY_BASE=5000000 source "<REPO_ROOT>/scripts/models/qwen3-235B-A22B.sh" && ray job submit
+MODEL_ARGS_LINE="$(MODEL_ARGS_ROTARY_BASE=5000000 python3 "<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py" qwen3-235B-A22B)" || exit 1; read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}" && ray job submit
   --address='http://127.0.0.1:8265'
   --runtime-env-json='{"env_vars": {"PYTHONUNBUFFERED": "1", "RAY_DEBUG": "1", "PYTHONPATH": "/root/Megatron-LM/", "CUDA_DEVICE_MAX_CONNECTIONS": "1", "NCCL_NVLS_ENABLE": "1", "MILES_LOG_DIR": "<SANDBOX>", "MC_TRANSFER_TIMEOUT": "300", "MODEL_ARGS_ROTARY_BASE": "5000000"}}'
   -- python3 "<REPO_ROOT>/train.py" ${MODEL_ARGS[@]}
diff --git a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Qwen3-235B-A22B-Instruct-2507/p2p.txt b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Qwen3-235B-A22B-Instruct-2507/p2p.txt
index 0a96a138158..274ec400820 100644
--- a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Qwen3-235B-A22B-Instruct-2507/p2p.txt
+++ b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Qwen3-235B-A22B-Instruct-2507/p2p.txt
@@ -39,7 +39,7 @@ RAY_memory_monitor_refresh_ms=0 ray start
 python3 -c "import ray; ray.init(address='auto', ignore_reinit_error=True); print(int(ray.cluster_resources().get('GPU', 0))); ray.shutdown()"
 
 ### 11
-MODEL_ARGS_ROTARY_BASE=5000000 source "<REPO_ROOT>/scripts/models/qwen3-235B-A22B.sh" && ray job submit
+MODEL_ARGS_LINE="$(MODEL_ARGS_ROTARY_BASE=5000000 python3 "<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py" qwen3-235B-A22B)" || exit 1; read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}" && ray job submit
   --address='http://127.0.0.1:8265'
   --runtime-env-json='{"env_vars": {"PYTHONUNBUFFERED": "1", "RAY_DEBUG": "1", "PYTHONPATH": "/root/Megatron-LM/", "CUDA_DEVICE_MAX_CONNECTIONS": "1", "NCCL_NVLS_ENABLE": "1", "MILES_LOG_DIR": "<SANDBOX>", "MC_TRANSFER_TIMEOUT": "300", "MODEL_ARGS_ROTARY_BASE": "5000000"}}'
   -- python3 "<REPO_ROOT>/train.py" ${MODEL_ARGS[@]}
diff --git a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Qwen3-30B-A3B/broadcast.txt b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Qwen3-30B-A3B/broadcast.txt
index 74ab1db0cac..1bcdc16346d 100644
--- a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Qwen3-30B-A3B/broadcast.txt
+++ b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Qwen3-30B-A3B/broadcast.txt
@@ -39,7 +39,7 @@ RAY_memory_monitor_refresh_ms=0 ray start
 python3 -c "import ray; ray.init(address='auto', ignore_reinit_error=True); print(int(ray.cluster_resources().get('GPU', 0))); ray.shutdown()"
 
 ### 11
-MODEL_ARGS_ROTARY_BASE=1000000 source "<REPO_ROOT>/scripts/models/qwen3-30B-A3B.sh" && ray job submit
+MODEL_ARGS_LINE="$(MODEL_ARGS_ROTARY_BASE=1000000 python3 "<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py" qwen3-30B-A3B)" || exit 1; read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}" && ray job submit
   --address='http://127.0.0.1:8265'
   --runtime-env-json='{"env_vars": {"PYTHONUNBUFFERED": "1", "RAY_DEBUG": "1", "PYTHONPATH": "/root/Megatron-LM/", "CUDA_DEVICE_MAX_CONNECTIONS": "1", "NCCL_NVLS_ENABLE": "1", "MILES_LOG_DIR": "<SANDBOX>", "MC_TRANSFER_TIMEOUT": "300", "MODEL_ARGS_ROTARY_BASE": "1000000"}}'
   -- python3 "<REPO_ROOT>/train.py" ${MODEL_ARGS[@]}
diff --git a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Qwen3-30B-A3B/p2p.txt b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Qwen3-30B-A3B/p2p.txt
index 259940107b6..a6e57018a74 100644
--- a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Qwen3-30B-A3B/p2p.txt
+++ b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Qwen3-30B-A3B/p2p.txt
@@ -39,7 +39,7 @@ RAY_memory_monitor_refresh_ms=0 ray start
 python3 -c "import ray; ray.init(address='auto', ignore_reinit_error=True); print(int(ray.cluster_resources().get('GPU', 0))); ray.shutdown()"
 
 ### 11
-MODEL_ARGS_ROTARY_BASE=1000000 source "<REPO_ROOT>/scripts/models/qwen3-30B-A3B.sh" && ray job submit
+MODEL_ARGS_LINE="$(MODEL_ARGS_ROTARY_BASE=1000000 python3 "<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py" qwen3-30B-A3B)" || exit 1; read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}" && ray job submit
   --address='http://127.0.0.1:8265'
   --runtime-env-json='{"env_vars": {"PYTHONUNBUFFERED": "1", "RAY_DEBUG": "1", "PYTHONPATH": "/root/Megatron-LM/", "CUDA_DEVICE_MAX_CONNECTIONS": "1", "NCCL_NVLS_ENABLE": "1", "MILES_LOG_DIR": "<SANDBOX>", "MC_TRANSFER_TIMEOUT": "300", "MODEL_ARGS_ROTARY_BASE": "1000000"}}'
   -- python3 "<REPO_ROOT>/train.py" ${MODEL_ARGS[@]}
diff --git a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Qwen3-4B/broadcast.txt b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Qwen3-4B/broadcast.txt
index 40fc6151875..5a8149a76ff 100644
--- a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Qwen3-4B/broadcast.txt
+++ b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Qwen3-4B/broadcast.txt
@@ -35,7 +35,7 @@ ray start
   --dashboard-port=8265
 
 ### 10
-source "<REPO_ROOT>/scripts/models/qwen3-4B.sh" && ray job submit
+MODEL_ARGS_LINE="$(python3 "<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py" qwen3-4B)" || exit 1; read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}" && ray job submit
   --address='http://127.0.0.1:8265'
   --runtime-env-json='{"env_vars": {"PYTHONUNBUFFERED": "1", "RAY_DEBUG": "1", "PYTHONPATH": "/root/Megatron-LM/", "CUDA_DEVICE_MAX_CONNECTIONS": "1", "NCCL_NVLS_ENABLE": "0", "MILES_LOG_DIR": "<SANDBOX>"}}'
   -- python3 "<REPO_ROOT>/train.py" ${MODEL_ARGS[@]}
diff --git a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Qwen3-4B/p2p.txt b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Qwen3-4B/p2p.txt
index c89b79aaf14..7480814ef17 100644
--- a/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Qwen3-4B/p2p.txt
+++ b/tests/snapshots/launch_scripts/self_executing/examples/infra_features/p2p_weight_transfer/run.py/run/Qwen3-4B/p2p.txt
@@ -35,7 +35,7 @@ ray start
   --dashboard-port=8265
 
 ### 10
-source "<REPO_ROOT>/scripts/models/qwen3-4B.sh" && ray job submit
+MODEL_ARGS_LINE="$(python3 "<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py" qwen3-4B)" || exit 1; read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}" && ray job submit
   --address='http://127.0.0.1:8265'
   --runtime-env-json='{"env_vars": {"PYTHONUNBUFFERED": "1", "RAY_DEBUG": "1", "PYTHONPATH": "/root/Megatron-LM/", "CUDA_DEVICE_MAX_CONNECTIONS": "1", "NCCL_NVLS_ENABLE": "0", "MILES_LOG_DIR": "<SANDBOX>"}}'
   -- python3 "<REPO_ROOT>/train.py" ${MODEL_ARGS[@]}
diff --git a/tests/snapshots/launch_scripts/sh/examples/experimental/eval/scripts/run-qwen3-32B.sh.txt b/tests/snapshots/launch_scripts/sh/examples/experimental/eval/scripts/run-qwen3-32B.sh.txt
index 4ce3827918b..1b4a112ce30 100644
--- a/tests/snapshots/launch_scripts/sh/examples/experimental/eval/scripts/run-qwen3-32B.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/experimental/eval/scripts/run-qwen3-32B.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py"
+"qwen3-32B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,7 +60,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/experimental/eval/scripts/run-qwen3-4B.sh.txt b/tests/snapshots/launch_scripts/sh/examples/experimental/eval/scripts/run-qwen3-4B.sh.txt
index 24bd9216b60..14392971a53 100644
--- a/tests/snapshots/launch_scripts/sh/examples/experimental/eval/scripts/run-qwen3-4B.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/experimental/eval/scripts/run-qwen3-4B.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py"
+"qwen3-4B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,7 +60,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/experimental/eval_multi_task/multi_task.sh.txt b/tests/snapshots/launch_scripts/sh/examples/experimental/eval_multi_task/multi_task.sh.txt
index 7217c79d1ff..a0a0c938015 100644
--- a/tests/snapshots/launch_scripts/sh/examples/experimental/eval_multi_task/multi_task.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/experimental/eval_multi_task/multi_task.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py"
+"qwen3-4B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,7 +60,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/experimental/multi_agent/run-qwen3-30B-A3B-multi-agent.sh.txt b/tests/snapshots/launch_scripts/sh/examples/experimental/multi_agent/run-qwen3-30B-A3B-multi-agent.sh.txt
index 797fee6c211..6cc7ba51340 100644
--- a/tests/snapshots/launch_scripts/sh/examples/experimental/multi_agent/run-qwen3-30B-A3B-multi-agent.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/experimental/multi_agent/run-qwen3-30B-A3B-multi-agent.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/examples/experimental/multi_agent/../../../miles/utils/external_utils/model_args_utils.py"
+"qwen3-30B-A3B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,7 +60,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/experimental/reproducibility/run-qwen2.5-0.5B-gsm8k.sh.txt b/tests/snapshots/launch_scripts/sh/examples/experimental/reproducibility/run-qwen2.5-0.5B-gsm8k.sh.txt
index 132b497801b..a10b60aa56a 100644
--- a/tests/snapshots/launch_scripts/sh/examples/experimental/reproducibility/run-qwen2.5-0.5B-gsm8k.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/experimental/reproducibility/run-qwen2.5-0.5B-gsm8k.sh.txt
@@ -39,6 +39,11 @@
 "python"
 
 ### 8
+"python3"
+"<REPO_ROOT>/examples/experimental/reproducibility/../../../miles/utils/external_utils/model_args_utils.py"
+"qwen2.5-0.5B"
+
+### 9
 "ray"
 "start"
 "--head"
@@ -48,7 +53,7 @@
 "8"
 "--disable-usage-stats"
 
-### 9
+### 10
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/experimental/search-r1/run_qwen2.5_3B.sh.txt b/tests/snapshots/launch_scripts/sh/examples/experimental/search-r1/run_qwen2.5_3B.sh.txt
index abbfec1ecad..88f140e55ed 100644
--- a/tests/snapshots/launch_scripts/sh/examples/experimental/search-r1/run_qwen2.5_3B.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/experimental/search-r1/run_qwen2.5_3B.sh.txt
@@ -39,6 +39,11 @@
 "python"
 
 ### 8
+"python3"
+"<REPO_ROOT>/examples/experimental/search-r1/../../../miles/utils/external_utils/model_args_utils.py"
+"qwen2.5-3B"
+
+### 9
 "ray"
 "start"
 "--head"
@@ -48,7 +53,7 @@
 "8"
 "--disable-usage-stats"
 
-### 9
+### 10
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/experimental/strands_sglang/strands_qwen3_8b.sh.txt b/tests/snapshots/launch_scripts/sh/examples/experimental/strands_sglang/strands_qwen3_8b.sh.txt
index b7344a73b12..b11aa347375 100644
--- a/tests/snapshots/launch_scripts/sh/examples/experimental/strands_sglang/strands_qwen3_8b.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/experimental/strands_sglang/strands_qwen3_8b.sh.txt
@@ -44,10 +44,15 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/examples/experimental/strands_sglang/../../../miles/utils/external_utils/model_args_utils.py"
+"qwen3-8B"
+
+### 10
 "date"
 "+%Y%m%d_%H%M%S"
 
-### 10
+### 11
 "ray"
 "start"
 "--head"
@@ -59,7 +64,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 11
+### 12
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/experimental/tau-bench/run_qwen3_4B.sh.txt b/tests/snapshots/launch_scripts/sh/examples/experimental/tau-bench/run_qwen3_4B.sh.txt
index d0fb68d3d7e..51b208db418 100644
--- a/tests/snapshots/launch_scripts/sh/examples/experimental/tau-bench/run_qwen3_4B.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/experimental/tau-bench/run_qwen3_4B.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/examples/experimental/tau-bench/../../../miles/utils/external_utils/model_args_utils.py"
+"qwen3-4B-Instruct-2507"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -57,7 +62,7 @@
 "--temp-dir"
 "/root/shared/ray_temp"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/fully_async/run-qwen3-4b-fully_async.sh.txt b/tests/snapshots/launch_scripts/sh/examples/fully_async/run-qwen3-4b-fully_async.sh.txt
index 4e35206c448..cf90b6930ee 100644
--- a/tests/snapshots/launch_scripts/sh/examples/fully_async/run-qwen3-4b-fully_async.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/fully_async/run-qwen3-4b-fully_async.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/examples/fully_async/../../miles/utils/external_utils/model_args_utils.py"
+"qwen3-4B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -53,7 +58,7 @@
 "8"
 "--disable-usage-stats"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/geo3k_vlm/run_geo3k_vlm.sh.txt b/tests/snapshots/launch_scripts/sh/examples/geo3k_vlm/run_geo3k_vlm.sh.txt
index e00be714548..dd9a072c6b1 100644
--- a/tests/snapshots/launch_scripts/sh/examples/geo3k_vlm/run_geo3k_vlm.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/geo3k_vlm/run_geo3k_vlm.sh.txt
@@ -71,6 +71,11 @@
 "/root/datasets/geo3k_imgurl"
 
 ### 13
+"python3"
+"<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py"
+"qwen3-8B"
+
+### 14
 "ray"
 "start"
 "--head"
@@ -82,7 +87,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 14
+### 15
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/geo3k_vlm/run_geo3k_vlm_sft.sh.txt b/tests/snapshots/launch_scripts/sh/examples/geo3k_vlm/run_geo3k_vlm_sft.sh.txt
index a2355b5e92e..55d09629604 100644
--- a/tests/snapshots/launch_scripts/sh/examples/geo3k_vlm/run_geo3k_vlm_sft.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/geo3k_vlm/run_geo3k_vlm_sft.sh.txt
@@ -71,6 +71,11 @@
 "/root/datasets/geo3k_imgurl"
 
 ### 13
+"python3"
+"<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py"
+"qwen3-8B"
+
+### 14
 "ray"
 "start"
 "--head"
@@ -82,7 +87,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 14
+### 15
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/infra_features/low_precision/run-kimi-k2-Thinking-int4.sh.txt b/tests/snapshots/launch_scripts/sh/examples/infra_features/low_precision/run-kimi-k2-Thinking-int4.sh.txt
index c32efbf09fd..cf1878c66a6 100644
--- a/tests/snapshots/launch_scripts/sh/examples/infra_features/low_precision/run-kimi-k2-Thinking-int4.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/infra_features/low_precision/run-kimi-k2-Thinking-int4.sh.txt
@@ -42,6 +42,11 @@
 "nvidia-smi"
 
 ### 9
+"python3"
+"<REPO_ROOT>/examples/infra_features/low_precision/../../../miles/utils/external_utils/model_args_utils.py"
+"kimi-k2-thinking"
+
+### 10
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/infra_features/low_precision/run-moonlight-16B-A3B-int4.sh.txt b/tests/snapshots/launch_scripts/sh/examples/infra_features/low_precision/run-moonlight-16B-A3B-int4.sh.txt
index 3c647619e0a..6df380bb002 100644
--- a/tests/snapshots/launch_scripts/sh/examples/infra_features/low_precision/run-moonlight-16B-A3B-int4.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/infra_features/low_precision/run-moonlight-16B-A3B-int4.sh.txt
@@ -49,6 +49,11 @@
 "-m"
 
 ### 10
+"python3"
+"<REPO_ROOT>/examples/infra_features/low_precision/../../../miles/utils/external_utils/model_args_utils.py"
+"moonlight"
+
+### 11
 "ray"
 "start"
 "--head"
@@ -60,7 +65,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 11
+### 12
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/infra_features/low_precision/run-qwen3-235B-A22B-int4.sh.txt b/tests/snapshots/launch_scripts/sh/examples/infra_features/low_precision/run-qwen3-235B-A22B-int4.sh.txt
index c5f34c7d39e..1b05aa7262f 100644
--- a/tests/snapshots/launch_scripts/sh/examples/infra_features/low_precision/run-qwen3-235B-A22B-int4.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/infra_features/low_precision/run-qwen3-235B-A22B-int4.sh.txt
@@ -42,6 +42,11 @@
 "nvidia-smi"
 
 ### 9
+"python3"
+"<REPO_ROOT>/examples/infra_features/low_precision/../../../miles/utils/external_utils/model_args_utils.py"
+"qwen3-235B-A22B"
+
+### 10
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/infra_features/low_precision/run-qwen3-30B-A3B-int4.sh.txt b/tests/snapshots/launch_scripts/sh/examples/infra_features/low_precision/run-qwen3-30B-A3B-int4.sh.txt
index d0b98417729..62fe888f50f 100644
--- a/tests/snapshots/launch_scripts/sh/examples/infra_features/low_precision/run-qwen3-30B-A3B-int4.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/infra_features/low_precision/run-qwen3-30B-A3B-int4.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/examples/infra_features/low_precision/../../../miles/utils/external_utils/model_args_utils.py"
+"qwen3-30B-A3B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,7 +60,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/infra_features/low_precision/run-qwen3-30b-a3b-fp8-two-nodes.sh.txt b/tests/snapshots/launch_scripts/sh/examples/infra_features/low_precision/run-qwen3-30b-a3b-fp8-two-nodes.sh.txt
index a89c5e24a69..7506aab4744 100644
--- a/tests/snapshots/launch_scripts/sh/examples/infra_features/low_precision/run-qwen3-30b-a3b-fp8-two-nodes.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/infra_features/low_precision/run-qwen3-30b-a3b-fp8-two-nodes.sh.txt
@@ -6,14 +6,19 @@
 "-m"
 
 ### 1
-"ps"
-"aux"
+"python3"
+"<REPO_ROOT>/examples/infra_features/low_precision/../../../miles/utils/external_utils/model_args_utils.py"
+"qwen3-30B-A3B"
 
 ### 2
 "ps"
 "aux"
 
 ### 3
+"ps"
+"aux"
+
+### 4
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/infra_features/low_precision/run-qwen3-4b-fp8.sh.txt b/tests/snapshots/launch_scripts/sh/examples/infra_features/low_precision/run-qwen3-4b-fp8.sh.txt
index 3580be18de6..3877f48f96c 100644
--- a/tests/snapshots/launch_scripts/sh/examples/infra_features/low_precision/run-qwen3-4b-fp8.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/infra_features/low_precision/run-qwen3-4b-fp8.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/examples/infra_features/low_precision/../../../miles/utils/external_utils/model_args_utils.py"
+"qwen3-4B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,7 +60,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/infra_features/p2p_weight_transfer/run-glm4.5-air-8node-profile.sh.txt b/tests/snapshots/launch_scripts/sh/examples/infra_features/p2p_weight_transfer/run-glm4.5-air-8node-profile.sh.txt
index 433a99f4522..a2f17d691da 100644
--- a/tests/snapshots/launch_scripts/sh/examples/infra_features/p2p_weight_transfer/run-glm4.5-air-8node-profile.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/infra_features/p2p_weight_transfer/run-glm4.5-air-8node-profile.sh.txt
@@ -45,10 +45,15 @@
 
 ### 9
 "python3"
+"<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py"
+"glm4.5-106B-A12B"
+
+### 10
+"python3"
 "-c"
 "print(int(1.0 * 1024 * 1024 * 1024))"
 
-### 10
+### 11
 "ray"
 "start"
 "--head"
@@ -60,22 +65,22 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 11
+### 12
 "python3"
 "-c"
 "import ray; ray.init(address='auto', ignore_reinit_error=True); print(int(ray.cluster_resources().get('GPU', 0))); ray.shutdown()"
 
-### 12
+### 13
 "mkdir"
 "-p"
 "<SANDBOX>/workdir"
 
-### 13
+### 14
 "rm"
 "-f"
 "<SANDBOX>/workdir/job_done_p2p"
 
-### 14
+### 15
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/infra_features/p2p_weight_transfer/run-glm4.7-flash-2node-profile.sh.txt b/tests/snapshots/launch_scripts/sh/examples/infra_features/p2p_weight_transfer/run-glm4.7-flash-2node-profile.sh.txt
index 2fcae95126e..033bacb2cc7 100644
--- a/tests/snapshots/launch_scripts/sh/examples/infra_features/p2p_weight_transfer/run-glm4.7-flash-2node-profile.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/infra_features/p2p_weight_transfer/run-glm4.7-flash-2node-profile.sh.txt
@@ -49,6 +49,11 @@
 "-m"
 
 ### 10
+"python3"
+"<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py"
+"glm4.7-flash"
+
+### 11
 "ray"
 "start"
 "--head"
@@ -60,12 +65,12 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 11
+### 12
 "python3"
 "-c"
 "import ray; ray.init(address='auto', ignore_reinit_error=True); print(int(ray.cluster_resources().get('GPU', 0))); ray.shutdown()"
 
-### 12
+### 13
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/infra_features/p2p_weight_transfer/run-glm5-disagg-profile.sh.txt b/tests/snapshots/launch_scripts/sh/examples/infra_features/p2p_weight_transfer/run-glm5-disagg-profile.sh.txt
index 03d6e0dd41c..23363db1791 100644
--- a/tests/snapshots/launch_scripts/sh/examples/infra_features/p2p_weight_transfer/run-glm5-disagg-profile.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/infra_features/p2p_weight_transfer/run-glm5-disagg-profile.sh.txt
@@ -44,6 +44,11 @@
 "redis"
 
 ### 9
+"python3"
+"<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py"
+"glm5-744B-A40B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,22 +60,22 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "python3"
 "-c"
 "import ray; ray.init(address='auto', ignore_reinit_error=True); print(int(ray.cluster_resources().get('GPU', 0))); ray.shutdown()"
 
-### 11
+### 12
 "mkdir"
 "-p"
 "<SANDBOX>/workdir"
 
-### 12
+### 13
 "rm"
 "-f"
 "<SANDBOX>/workdir/job_done_p2p"
 
-### 13
+### 14
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/infra_features/p2p_weight_transfer/run-kimi-k2-64node-profile.sh.txt b/tests/snapshots/launch_scripts/sh/examples/infra_features/p2p_weight_transfer/run-kimi-k2-64node-profile.sh.txt
index b53d6ae242a..724c1f63b9c 100644
--- a/tests/snapshots/launch_scripts/sh/examples/infra_features/p2p_weight_transfer/run-kimi-k2-64node-profile.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/infra_features/p2p_weight_transfer/run-kimi-k2-64node-profile.sh.txt
@@ -49,6 +49,11 @@
 "-m"
 
 ### 10
+"python3"
+"<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py"
+"kimi-k2"
+
+### 11
 "ray"
 "start"
 "--head"
@@ -60,12 +65,12 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 11
+### 12
 "python3"
 "-c"
 "import ray; ray.init(address='auto', ignore_reinit_error=True); print(int(ray.cluster_resources().get('GPU', 0))); ray.shutdown()"
 
-### 12
+### 13
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/infra_features/p2p_weight_transfer/run-qwen3-235B-A22B-16node-profile.sh.txt b/tests/snapshots/launch_scripts/sh/examples/infra_features/p2p_weight_transfer/run-qwen3-235B-A22B-16node-profile.sh.txt
index 6576a46dad4..36d913e8721 100644
--- a/tests/snapshots/launch_scripts/sh/examples/infra_features/p2p_weight_transfer/run-qwen3-235B-A22B-16node-profile.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/infra_features/p2p_weight_transfer/run-qwen3-235B-A22B-16node-profile.sh.txt
@@ -50,10 +50,15 @@
 
 ### 10
 "python3"
+"<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py"
+"qwen3-235B-A22B"
+
+### 11
+"python3"
 "-c"
 "print(int(1.0 * 1024 * 1024 * 1024))"
 
-### 11
+### 12
 "ray"
 "start"
 "--head"
@@ -65,12 +70,12 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 12
+### 13
 "python3"
 "-c"
 "import ray; ray.init(address='auto', ignore_reinit_error=True); print(int(ray.cluster_resources().get('GPU', 0))); ray.shutdown()"
 
-### 13
+### 14
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/infra_features/p2p_weight_transfer/run-qwen3-30B-A3B-4node-profile.sh.txt b/tests/snapshots/launch_scripts/sh/examples/infra_features/p2p_weight_transfer/run-qwen3-30B-A3B-4node-profile.sh.txt
index d04fe971f32..42937904913 100644
--- a/tests/snapshots/launch_scripts/sh/examples/infra_features/p2p_weight_transfer/run-qwen3-30B-A3B-4node-profile.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/infra_features/p2p_weight_transfer/run-qwen3-30B-A3B-4node-profile.sh.txt
@@ -49,6 +49,11 @@
 "-m"
 
 ### 10
+"python3"
+"<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py"
+"qwen3-30B-A3B"
+
+### 11
 "ray"
 "start"
 "--head"
@@ -60,12 +65,12 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 11
+### 12
 "python3"
 "-c"
 "import ray; ray.init(address='auto', ignore_reinit_error=True); print(int(ray.cluster_resources().get('GPU', 0))); ray.shutdown()"
 
-### 12
+### 13
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/infra_features/train_infer_mismatch_helper/run-qwen3-4b-mis.sh.txt b/tests/snapshots/launch_scripts/sh/examples/infra_features/train_infer_mismatch_helper/run-qwen3-4b-mis.sh.txt
index 5d0cf0562d9..0b47754d4c5 100644
--- a/tests/snapshots/launch_scripts/sh/examples/infra_features/train_infer_mismatch_helper/run-qwen3-4b-mis.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/infra_features/train_infer_mismatch_helper/run-qwen3-4b-mis.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/examples/infra_features/train_infer_mismatch_helper/../../../miles/utils/external_utils/model_args_utils.py"
+"qwen3-4B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,7 +60,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/lora/dev.sh.txt b/tests/snapshots/launch_scripts/sh/examples/lora/dev.sh.txt
index dda91e58a86..d23df4cf680 100644
--- a/tests/snapshots/launch_scripts/sh/examples/lora/dev.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/lora/dev.sh.txt
@@ -29,6 +29,11 @@
 "python"
 
 ### 6
+"python3"
+"<REPO_ROOT>/examples/lora/../../miles/utils/external_utils/model_args_utils.py"
+"qwen2.5-3B"
+
+### 7
 "ray"
 "start"
 "--head"
@@ -38,7 +43,7 @@
 "1"
 "--disable-usage-stats"
 
-### 7
+### 8
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/lora/run-gpt-oss-20B-megatron-moe-lora.sh.txt b/tests/snapshots/launch_scripts/sh/examples/lora/run-gpt-oss-20B-megatron-moe-lora.sh.txt
index a1025b6d466..cfc63b109a2 100644
--- a/tests/snapshots/launch_scripts/sh/examples/lora/run-gpt-oss-20B-megatron-moe-lora.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/lora/run-gpt-oss-20B-megatron-moe-lora.sh.txt
@@ -39,6 +39,11 @@
 "python"
 
 ### 8
+"python3"
+"<REPO_ROOT>/examples/lora/../../miles/utils/external_utils/model_args_utils.py"
+"gpt-oss-20b"
+
+### 9
 "ray"
 "start"
 "--head"
@@ -50,7 +55,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 9
+### 10
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/lora/run-kimi-k25-megatron-lora.sh.txt b/tests/snapshots/launch_scripts/sh/examples/lora/run-kimi-k25-megatron-lora.sh.txt
index 48d1e0ee363..a92c64589ef 100644
--- a/tests/snapshots/launch_scripts/sh/examples/lora/run-kimi-k25-megatron-lora.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/lora/run-kimi-k25-megatron-lora.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/examples/lora/../../miles/utils/external_utils/model_args_utils.py"
+"kimi-k2-thinking"
+
+### 10
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/lora/run-qwen2.5-0.5B-megatron-lora.sh.txt b/tests/snapshots/launch_scripts/sh/examples/lora/run-qwen2.5-0.5B-megatron-lora.sh.txt
index 87c29860d42..8ffdd1560aa 100644
--- a/tests/snapshots/launch_scripts/sh/examples/lora/run-qwen2.5-0.5B-megatron-lora.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/lora/run-qwen2.5-0.5B-megatron-lora.sh.txt
@@ -29,6 +29,11 @@
 "python"
 
 ### 6
+"python3"
+"<REPO_ROOT>/examples/lora/../../miles/utils/external_utils/model_args_utils.py"
+"qwen2.5-0.5B"
+
+### 7
 "ray"
 "start"
 "--head"
@@ -38,7 +43,7 @@
 "8"
 "--disable-usage-stats"
 
-### 7
+### 8
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/lora/run-qwen2.5-3B-megatron-lora-disaggregated-multi-node.sh.txt b/tests/snapshots/launch_scripts/sh/examples/lora/run-qwen2.5-3B-megatron-lora-disaggregated-multi-node.sh.txt
index a0c726c5143..69deec869e8 100644
--- a/tests/snapshots/launch_scripts/sh/examples/lora/run-qwen2.5-3B-megatron-lora-disaggregated-multi-node.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/lora/run-qwen2.5-3B-megatron-lora-disaggregated-multi-node.sh.txt
@@ -53,6 +53,11 @@
 "python"
 
 ### 10
+"python3"
+"<REPO_ROOT>/examples/lora/../../miles/utils/external_utils/model_args_utils.py"
+"qwen2.5-3B"
+
+### 11
 "ray"
 "start"
 "--head"
@@ -64,12 +69,12 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 11
+### 12
 "python3"
 "-c"
 "import ray; ray.init(address='auto', ignore_reinit_error=True); print(int(ray.cluster_resources().get('GPU', 0))); ray.shutdown()"
 
-### 12
+### 13
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/lora/run-qwen2.5-3B-megatron-lora-disaggregated.sh.txt b/tests/snapshots/launch_scripts/sh/examples/lora/run-qwen2.5-3B-megatron-lora-disaggregated.sh.txt
index f6c62b12adc..ef8dee4b7ae 100644
--- a/tests/snapshots/launch_scripts/sh/examples/lora/run-qwen2.5-3B-megatron-lora-disaggregated.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/lora/run-qwen2.5-3B-megatron-lora-disaggregated.sh.txt
@@ -29,6 +29,11 @@
 "python"
 
 ### 6
+"python3"
+"<REPO_ROOT>/examples/lora/../../miles/utils/external_utils/model_args_utils.py"
+"qwen2.5-3B"
+
+### 7
 "ray"
 "start"
 "--head"
@@ -38,7 +43,7 @@
 "2"
 "--disable-usage-stats"
 
-### 7
+### 8
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/lora/run-qwen3-4B-megatron-lora.sh.txt b/tests/snapshots/launch_scripts/sh/examples/lora/run-qwen3-4B-megatron-lora.sh.txt
index e70fe24e1c3..6864acd787d 100644
--- a/tests/snapshots/launch_scripts/sh/examples/lora/run-qwen3-4B-megatron-lora.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/lora/run-qwen3-4B-megatron-lora.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/miles/utils/external_utils/model_args_utils.py"
+"qwen3-4B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,7 +60,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/lora/run-qwen3-4b-megatron-lora-result.sh.txt b/tests/snapshots/launch_scripts/sh/examples/lora/run-qwen3-4b-megatron-lora-result.sh.txt
index c7444461f45..82a54662016 100644
--- a/tests/snapshots/launch_scripts/sh/examples/lora/run-qwen3-4b-megatron-lora-result.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/lora/run-qwen3-4b-megatron-lora-result.sh.txt
@@ -42,6 +42,11 @@
 "nvidia-smi"
 
 ### 9
+"python3"
+"<REPO_ROOT>/examples/lora/../../miles/utils/external_utils/model_args_utils.py"
+"qwen3-4B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -51,7 +56,7 @@
 "4"
 "--disable-usage-stats"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/examples/on_policy_distillation/run-qwen3-8B-opd-megatron.sh.txt b/tests/snapshots/launch_scripts/sh/examples/on_policy_distillation/run-qwen3-8B-opd-megatron.sh.txt
index d4a648418ab..d746325862c 100644
--- a/tests/snapshots/launch_scripts/sh/examples/on_policy_distillation/run-qwen3-8B-opd-megatron.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/on_policy_distillation/run-qwen3-8B-opd-megatron.sh.txt
@@ -6,6 +6,11 @@
 "-m"
 
 ### 1
+"python3"
+"<REPO_ROOT>/examples/on_policy_distillation/../../miles/utils/external_utils/model_args_utils.py"
+"qwen3-8B"
+
+### 2
 "ray"
 "start"
 "--head"
@@ -17,7 +22,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 2
+### 3
 "ray"
 "job"
 "submit"
@@ -150,40 +155,40 @@
 "--rm-type"
 "math"
 
-### 3
+### 4
 "pkill"
 "-9"
 "sglang"
 
-### 4
+### 5
 "sleep"
 "3"
 
-### 5
+### 6
 "ray"
 "stop"
 "--force"
 
-### 6
+### 7
 "pkill"
 "-9"
 "ray"
 
-### 7
+### 8
 "pkill"
 "-9"
 "python"
 
-### 8
+### 9
 "sleep"
 "3"
 
-### 9
+### 10
 "pkill"
 "-9"
 "ray"
 
-### 10
+### 11
 "pkill"
 "-9"
 "python"
diff --git a/tests/snapshots/launch_scripts/sh/examples/on_policy_distillation/run-qwen3-8B-opd-multi-teacher.sh.txt b/tests/snapshots/launch_scripts/sh/examples/on_policy_distillation/run-qwen3-8B-opd-multi-teacher.sh.txt
index 590060a19f6..2ee5f94badc 100644
--- a/tests/snapshots/launch_scripts/sh/examples/on_policy_distillation/run-qwen3-8B-opd-multi-teacher.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/on_policy_distillation/run-qwen3-8B-opd-multi-teacher.sh.txt
@@ -71,6 +71,11 @@
 "-m"
 
 ### 10
+"python3"
+"<REPO_ROOT>/examples/on_policy_distillation/../../miles/utils/external_utils/model_args_utils.py"
+"qwen3-8B"
+
+### 11
 "ray"
 "start"
 "--head"
@@ -82,7 +87,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 11
+### 12
 "ray"
 "job"
 "submit"
@@ -227,40 +232,40 @@
 "--opd-teacher-key"
 "opd_teacher"
 
-### 12
+### 13
 "pkill"
 "-9"
 "sglang"
 
-### 13
+### 14
 "sleep"
 "3"
 
-### 14
+### 15
 "ray"
 "stop"
 "--force"
 
-### 15
+### 16
 "pkill"
 "-9"
 "ray"
 
-### 16
+### 17
 "pkill"
 "-9"
 "python"
 
-### 17
+### 18
 "sleep"
 "3"
 
-### 18
+### 19
 "pkill"
 "-9"
 "ray"
 
-### 19
+### 20
 "pkill"
 "-9"
 "python"
diff --git a/tests/snapshots/launch_scripts/sh/examples/on_policy_distillation/run-qwen3-8B-opd.sh.txt b/tests/snapshots/launch_scripts/sh/examples/on_policy_distillation/run-qwen3-8B-opd.sh.txt
index 7e32ebd59d6..33b048a20a4 100644
--- a/tests/snapshots/launch_scripts/sh/examples/on_policy_distillation/run-qwen3-8B-opd.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/examples/on_policy_distillation/run-qwen3-8B-opd.sh.txt
@@ -36,6 +36,11 @@
 "-m"
 
 ### 5
+"python3"
+"<REPO_ROOT>/examples/on_policy_distillation/../../miles/utils/external_utils/model_args_utils.py"
+"qwen3-8B"
+
+### 6
 "ray"
 "start"
 "--head"
@@ -47,7 +52,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 6
+### 7
 "ray"
 "job"
 "submit"
@@ -188,40 +193,40 @@
 "--rm-url"
 "http://127.0.0.1:13141/generate"
 
-### 7
+### 8
 "pkill"
 "-9"
 "sglang"
 
-### 8
+### 9
 "sleep"
 "3"
 
-### 9
+### 10
 "ray"
 "stop"
 "--force"
 
-### 10
+### 11
 "pkill"
 "-9"
 "ray"
 
-### 11
+### 12
 "pkill"
 "-9"
 "python"
 
-### 12
+### 13
 "sleep"
 "3"
 
-### 13
+### 14
 "pkill"
 "-9"
 "ray"
 
-### 14
+### 15
 "pkill"
 "-9"
 "python"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/amd/run-qwen3-4B-amd.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/amd/run-qwen3-4B-amd.sh.txt
index 05e1b71c70a..553ca0e6315 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/amd/run-qwen3-4B-amd.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/amd/run-qwen3-4B-amd.sh.txt
@@ -39,6 +39,11 @@
 "python"
 
 ### 8
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"qwen3-4B"
+
+### 9
 "ray"
 "start"
 "--head"
@@ -50,7 +55,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 9
+### 10
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-deepseek-r1.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-deepseek-r1.sh.txt
index c3bf834fc9e..ee964dfc2f3 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-deepseek-r1.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-deepseek-r1.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"deepseek-v3"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,7 +60,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-glm4-9B-4xgpu-radixtree.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-glm4-9B-4xgpu-radixtree.sh.txt
index 9e5b0b4f7f0..a6e70ac09f7 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-glm4-9B-4xgpu-radixtree.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-glm4-9B-4xgpu-radixtree.sh.txt
@@ -42,6 +42,11 @@
 "nvidia-smi"
 
 ### 9
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"glm4-9B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -53,7 +58,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-glm4-9B.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-glm4-9B.sh.txt
index 26f5eddcd26..36fe469186c 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-glm4-9B.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-glm4-9B.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"glm4-9B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,7 +60,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-glm4.5-355B-A32B.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-glm4.5-355B-A32B.sh.txt
index fb0cb552353..c7af53a0b7c 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-glm4.5-355B-A32B.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-glm4.5-355B-A32B.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"glm4.5-355B-A32B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -54,12 +59,12 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "awk"
 "{print $1}"
 "/root/mpi_rack_hostfile"
 
-### 11
+### 12
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-glm4.7-flash.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-glm4.7-flash.sh.txt
index b1fd672585e..9262d58117d 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-glm4.7-flash.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-glm4.7-flash.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"glm4.7-flash"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,7 +60,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-gpt-oss-20b-bf16.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-gpt-oss-20b-bf16.sh.txt
index 577a4b430c2..093da50e7a6 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-gpt-oss-20b-bf16.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-gpt-oss-20b-bf16.sh.txt
@@ -39,6 +39,11 @@
 "python"
 
 ### 8
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"gpt-oss-20b"
+
+### 9
 "ray"
 "start"
 "--head"
@@ -50,7 +55,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 9
+### 10
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-kimi-k2-Instruct.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-kimi-k2-Instruct.sh.txt
index 66dad480f75..11a38513a36 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-kimi-k2-Instruct.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-kimi-k2-Instruct.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"kimi-k2"
+
+### 10
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-kimi-k2-Thinking.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-kimi-k2-Thinking.sh.txt
index 88a8cd890fc..fc6e3a6def9 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-kimi-k2-Thinking.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-kimi-k2-Thinking.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"kimi-k2-thinking"
+
+### 10
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-kimi-k25.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-kimi-k25.sh.txt
index 909108edc3b..47cb8d8e819 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-kimi-k25.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-kimi-k25.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"kimi-k2-thinking"
+
+### 10
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-mimo-7B-rl-eagle.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-mimo-7B-rl-eagle.sh.txt
index b8b62f49951..9fa1648672f 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-mimo-7B-rl-eagle.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-mimo-7B-rl-eagle.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"mimo-7B-rl"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -53,7 +58,7 @@
 "8"
 "--disable-usage-stats"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-moonlight-16B-A3B.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-moonlight-16B-A3B.sh.txt
index b71b7296deb..278c7479a23 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-moonlight-16B-A3B.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-moonlight-16B-A3B.sh.txt
@@ -49,6 +49,11 @@
 "-m"
 
 ### 10
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"moonlight"
+
+### 11
 "ray"
 "start"
 "--head"
@@ -60,7 +65,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 11
+### 12
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-nemotron-3-nano-30b-a3b.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-nemotron-3-nano-30b-a3b.sh.txt
index 3e128f7144b..fa7ca0d683d 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-nemotron-3-nano-30b-a3b.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-nemotron-3-nano-30b-a3b.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"nemotron-3-nano-30b-a3b"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,7 +60,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-nemotron-3-nano-4b.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-nemotron-3-nano-4b.sh.txt
index 25060e59960..908fb643bb9 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-nemotron-3-nano-4b.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-nemotron-3-nano-4b.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"nemotron-3-nano-4b"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,7 +60,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-nemotron-3-super-120b-a12b.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-nemotron-3-super-120b-a12b.sh.txt
index 97010455b1c..d19fef80bc5 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-nemotron-3-super-120b-a12b.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-nemotron-3-super-120b-a12b.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"nemotron-3-super-120b-a12b"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,971 +60,971 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "ray"
 "status"
 
-### 11
+### 12
 "sleep"
 "5"
 
-### 12
+### 13
 "ray"
 "status"
 
-### 13
+### 14
 "sleep"
 "5"
 
-### 14
+### 15
 "ray"
 "status"
 
-### 15
+### 16
 "sleep"
 "5"
 
-### 16
+### 17
 "ray"
 "status"
 
-### 17
+### 18
 "sleep"
 "5"
 
-### 18
+### 19
 "ray"
 "status"
 
-### 19
+### 20
 "sleep"
 "5"
 
-### 20
+### 21
 "ray"
 "status"
 
-### 21
+### 22
 "sleep"
 "5"
 
-### 22
+### 23
 "ray"
 "status"
 
-### 23
+### 24
 "sleep"
 "5"
 
-### 24
+### 25
 "ray"
 "status"
 
-### 25
+### 26
 "sleep"
 "5"
 
-### 26
+### 27
 "ray"
 "status"
 
-### 27
+### 28
 "sleep"
 "5"
 
-### 28
+### 29
 "ray"
 "status"
 
-### 29
+### 30
 "sleep"
 "5"
 
-### 30
+### 31
 "ray"
 "status"
 
-### 31
+### 32
 "sleep"
 "5"
 
-### 32
+### 33
 "ray"
 "status"
 
-### 33
+### 34
 "sleep"
 "5"
 
-### 34
+### 35
 "ray"
 "status"
 
-### 35
+### 36
 "sleep"
 "5"
 
-### 36
+### 37
 "ray"
 "status"
 
-### 37
+### 38
 "sleep"
 "5"
 
-### 38
+### 39
 "ray"
 "status"
 
-### 39
+### 40
 "sleep"
 "5"
 
-### 40
+### 41
 "ray"
 "status"
 
-### 41
+### 42
 "sleep"
 "5"
 
-### 42
+### 43
 "ray"
 "status"
 
-### 43
+### 44
 "sleep"
 "5"
 
-### 44
+### 45
 "ray"
 "status"
 
-### 45
+### 46
 "sleep"
 "5"
 
-### 46
+### 47
 "ray"
 "status"
 
-### 47
+### 48
 "sleep"
 "5"
 
-### 48
+### 49
 "ray"
 "status"
 
-### 49
+### 50
 "sleep"
 "5"
 
-### 50
+### 51
 "ray"
 "status"
 
-### 51
+### 52
 "sleep"
 "5"
 
-### 52
+### 53
 "ray"
 "status"
 
-### 53
+### 54
 "sleep"
 "5"
 
-### 54
+### 55
 "ray"
 "status"
 
-### 55
+### 56
 "sleep"
 "5"
 
-### 56
+### 57
 "ray"
 "status"
 
-### 57
+### 58
 "sleep"
 "5"
 
-### 58
+### 59
 "ray"
 "status"
 
-### 59
+### 60
 "sleep"
 "5"
 
-### 60
+### 61
 "ray"
 "status"
 
-### 61
+### 62
 "sleep"
 "5"
 
-### 62
+### 63
 "ray"
 "status"
 
-### 63
+### 64
 "sleep"
 "5"
 
-### 64
+### 65
 "ray"
 "status"
 
-### 65
+### 66
 "sleep"
 "5"
 
-### 66
+### 67
 "ray"
 "status"
 
-### 67
+### 68
 "sleep"
 "5"
 
-### 68
+### 69
 "ray"
 "status"
 
-### 69
+### 70
 "sleep"
 "5"
 
-### 70
+### 71
 "ray"
 "status"
 
-### 71
+### 72
 "sleep"
 "5"
 
-### 72
+### 73
 "ray"
 "status"
 
-### 73
+### 74
 "sleep"
 "5"
 
-### 74
+### 75
 "ray"
 "status"
 
-### 75
+### 76
 "sleep"
 "5"
 
-### 76
+### 77
 "ray"
 "status"
 
-### 77
+### 78
 "sleep"
 "5"
 
-### 78
+### 79
 "ray"
 "status"
 
-### 79
+### 80
 "sleep"
 "5"
 
-### 80
+### 81
 "ray"
 "status"
 
-### 81
+### 82
 "sleep"
 "5"
 
-### 82
+### 83
 "ray"
 "status"
 
-### 83
+### 84
 "sleep"
 "5"
 
-### 84
+### 85
 "ray"
 "status"
 
-### 85
+### 86
 "sleep"
 "5"
 
-### 86
+### 87
 "ray"
 "status"
 
-### 87
+### 88
 "sleep"
 "5"
 
-### 88
+### 89
 "ray"
 "status"
 
-### 89
+### 90
 "sleep"
 "5"
 
-### 90
+### 91
 "ray"
 "status"
 
-### 91
+### 92
 "sleep"
 "5"
 
-### 92
+### 93
 "ray"
 "status"
 
-### 93
+### 94
 "sleep"
 "5"
 
-### 94
+### 95
 "ray"
 "status"
 
-### 95
+### 96
 "sleep"
 "5"
 
-### 96
+### 97
 "ray"
 "status"
 
-### 97
+### 98
 "sleep"
 "5"
 
-### 98
+### 99
 "ray"
 "status"
 
-### 99
+### 100
 "sleep"
 "5"
 
-### 100
+### 101
 "ray"
 "status"
 
-### 101
+### 102
 "sleep"
 "5"
 
-### 102
+### 103
 "ray"
 "status"
 
-### 103
+### 104
 "sleep"
 "5"
 
-### 104
+### 105
 "ray"
 "status"
 
-### 105
+### 106
 "sleep"
 "5"
 
-### 106
+### 107
 "ray"
 "status"
 
-### 107
+### 108
 "sleep"
 "5"
 
-### 108
+### 109
 "ray"
 "status"
 
-### 109
+### 110
 "sleep"
 "5"
 
-### 110
+### 111
 "ray"
 "status"
 
-### 111
+### 112
 "sleep"
 "5"
 
-### 112
+### 113
 "ray"
 "status"
 
-### 113
+### 114
 "sleep"
 "5"
 
-### 114
+### 115
 "ray"
 "status"
 
-### 115
+### 116
 "sleep"
 "5"
 
-### 116
+### 117
 "ray"
 "status"
 
-### 117
+### 118
 "sleep"
 "5"
 
-### 118
+### 119
 "ray"
 "status"
 
-### 119
+### 120
 "sleep"
 "5"
 
-### 120
+### 121
 "ray"
 "status"
 
-### 121
+### 122
 "sleep"
 "5"
 
-### 122
+### 123
 "ray"
 "status"
 
-### 123
+### 124
 "sleep"
 "5"
 
-### 124
+### 125
 "ray"
 "status"
 
-### 125
+### 126
 "sleep"
 "5"
 
-### 126
+### 127
 "ray"
 "status"
 
-### 127
+### 128
 "sleep"
 "5"
 
-### 128
+### 129
 "ray"
 "status"
 
-### 129
+### 130
 "sleep"
 "5"
 
-### 130
+### 131
 "ray"
 "status"
 
-### 131
+### 132
 "sleep"
 "5"
 
-### 132
+### 133
 "ray"
 "status"
 
-### 133
+### 134
 "sleep"
 "5"
 
-### 134
+### 135
 "ray"
 "status"
 
-### 135
+### 136
 "sleep"
 "5"
 
-### 136
+### 137
 "ray"
 "status"
 
-### 137
+### 138
 "sleep"
 "5"
 
-### 138
+### 139
 "ray"
 "status"
 
-### 139
+### 140
 "sleep"
 "5"
 
-### 140
+### 141
 "ray"
 "status"
 
-### 141
+### 142
 "sleep"
 "5"
 
-### 142
+### 143
 "ray"
 "status"
 
-### 143
+### 144
 "sleep"
 "5"
 
-### 144
+### 145
 "ray"
 "status"
 
-### 145
+### 146
 "sleep"
 "5"
 
-### 146
+### 147
 "ray"
 "status"
 
-### 147
+### 148
 "sleep"
 "5"
 
-### 148
+### 149
 "ray"
 "status"
 
-### 149
+### 150
 "sleep"
 "5"
 
-### 150
+### 151
 "ray"
 "status"
 
-### 151
+### 152
 "sleep"
 "5"
 
-### 152
+### 153
 "ray"
 "status"
 
-### 153
+### 154
 "sleep"
 "5"
 
-### 154
+### 155
 "ray"
 "status"
 
-### 155
+### 156
 "sleep"
 "5"
 
-### 156
+### 157
 "ray"
 "status"
 
-### 157
+### 158
 "sleep"
 "5"
 
-### 158
+### 159
 "ray"
 "status"
 
-### 159
+### 160
 "sleep"
 "5"
 
-### 160
+### 161
 "ray"
 "status"
 
-### 161
+### 162
 "sleep"
 "5"
 
-### 162
+### 163
 "ray"
 "status"
 
-### 163
+### 164
 "sleep"
 "5"
 
-### 164
+### 165
 "ray"
 "status"
 
-### 165
+### 166
 "sleep"
 "5"
 
-### 166
+### 167
 "ray"
 "status"
 
-### 167
+### 168
 "sleep"
 "5"
 
-### 168
+### 169
 "ray"
 "status"
 
-### 169
+### 170
 "sleep"
 "5"
 
-### 170
+### 171
 "ray"
 "status"
 
-### 171
+### 172
 "sleep"
 "5"
 
-### 172
+### 173
 "ray"
 "status"
 
-### 173
+### 174
 "sleep"
 "5"
 
-### 174
+### 175
 "ray"
 "status"
 
-### 175
+### 176
 "sleep"
 "5"
 
-### 176
+### 177
 "ray"
 "status"
 
-### 177
+### 178
 "sleep"
 "5"
 
-### 178
+### 179
 "ray"
 "status"
 
-### 179
+### 180
 "sleep"
 "5"
 
-### 180
+### 181
 "ray"
 "status"
 
-### 181
+### 182
 "sleep"
 "5"
 
-### 182
+### 183
 "ray"
 "status"
 
-### 183
+### 184
 "sleep"
 "5"
 
-### 184
+### 185
 "ray"
 "status"
 
-### 185
+### 186
 "sleep"
 "5"
 
-### 186
+### 187
 "ray"
 "status"
 
-### 187
+### 188
 "sleep"
 "5"
 
-### 188
+### 189
 "ray"
 "status"
 
-### 189
+### 190
 "sleep"
 "5"
 
-### 190
+### 191
 "ray"
 "status"
 
-### 191
+### 192
 "sleep"
 "5"
 
-### 192
+### 193
 "ray"
 "status"
 
-### 193
+### 194
 "sleep"
 "5"
 
-### 194
+### 195
 "ray"
 "status"
 
-### 195
+### 196
 "sleep"
 "5"
 
-### 196
+### 197
 "ray"
 "status"
 
-### 197
+### 198
 "sleep"
 "5"
 
-### 198
+### 199
 "ray"
 "status"
 
-### 199
+### 200
 "sleep"
 "5"
 
-### 200
+### 201
 "ray"
 "status"
 
-### 201
+### 202
 "sleep"
 "5"
 
-### 202
+### 203
 "ray"
 "status"
 
-### 203
+### 204
 "sleep"
 "5"
 
-### 204
+### 205
 "ray"
 "status"
 
-### 205
+### 206
 "sleep"
 "5"
 
-### 206
+### 207
 "ray"
 "status"
 
-### 207
+### 208
 "sleep"
 "5"
 
-### 208
+### 209
 "ray"
 "status"
 
-### 209
+### 210
 "sleep"
 "5"
 
-### 210
+### 211
 "ray"
 "status"
 
-### 211
+### 212
 "sleep"
 "5"
 
-### 212
+### 213
 "ray"
 "status"
 
-### 213
+### 214
 "sleep"
 "5"
 
-### 214
+### 215
 "ray"
 "status"
 
-### 215
+### 216
 "sleep"
 "5"
 
-### 216
+### 217
 "ray"
 "status"
 
-### 217
+### 218
 "sleep"
 "5"
 
-### 218
+### 219
 "ray"
 "status"
 
-### 219
+### 220
 "sleep"
 "5"
 
-### 220
+### 221
 "ray"
 "status"
 
-### 221
+### 222
 "sleep"
 "5"
 
-### 222
+### 223
 "ray"
 "status"
 
-### 223
+### 224
 "sleep"
 "5"
 
-### 224
+### 225
 "ray"
 "status"
 
-### 225
+### 226
 "sleep"
 "5"
 
-### 226
+### 227
 "ray"
 "status"
 
-### 227
+### 228
 "sleep"
 "5"
 
-### 228
+### 229
 "ray"
 "status"
 
-### 229
+### 230
 "sleep"
 "5"
 
-### 230
+### 231
 "ray"
 "status"
 
-### 231
+### 232
 "sleep"
 "5"
 
-### 232
+### 233
 "ray"
 "status"
 
-### 233
+### 234
 "sleep"
 "5"
 
-### 234
+### 235
 "ray"
 "status"
 
-### 235
+### 236
 "sleep"
 "5"
 
-### 236
+### 237
 "ray"
 "status"
 
-### 237
+### 238
 "sleep"
 "5"
 
-### 238
+### 239
 "ray"
 "status"
 
-### 239
+### 240
 "sleep"
 "5"
 
-### 240
+### 241
 "ray"
 "status"
 
-### 241
+### 242
 "sleep"
 "5"
 
-### 242
+### 243
 "ray"
 "status"
 
-### 243
+### 244
 "sleep"
 "5"
 
-### 244
+### 245
 "ray"
 "status"
 
-### 245
+### 246
 "sleep"
 "5"
 
-### 246
+### 247
 "ray"
 "status"
 
-### 247
+### 248
 "sleep"
 "5"
 
-### 248
+### 249
 "ray"
 "status"
 
-### 249
+### 250
 "sleep"
 "5"
 
-### 250
+### 251
 "ray"
 "status"
 
-### 251
+### 252
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-235B-A22B-sft.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-235B-A22B-sft.sh.txt
index e91727a97db..3915b2df14f 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-235B-A22B-sft.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-235B-A22B-sft.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"qwen3-235B-A22B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,12 +60,12 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "awk"
 "{print $1}"
 "/root/mpi_rack_hostfile"
 
-### 11
+### 12
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-235B-A22B.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-235B-A22B.sh.txt
index 2387b134bae..801704f8d08 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-235B-A22B.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-235B-A22B.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"qwen3-235B-A22B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,12 +60,12 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "awk"
 "{print $1}"
 "/root/mpi_rack_hostfile"
 
-### 11
+### 12
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-32B.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-32B.sh.txt
index 9fab9236ead..8adb2d846f7 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-32B.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-32B.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"qwen3-32B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,7 +60,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-4B-base-sft.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-4B-base-sft.sh.txt
index 3563312a8c8..de7ee3f9dc6 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-4B-base-sft.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-4B-base-sft.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"qwen3-4B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,7 +60,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-4B.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-4B.sh.txt
index d4c1bb37af7..62832339fc2 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-4B.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-4B.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"qwen3-4B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,7 +60,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-4B_4xgpu.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-4B_4xgpu.sh.txt
index 5aca51b031f..0ac3ee9d6ad 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-4B_4xgpu.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-4B_4xgpu.sh.txt
@@ -42,6 +42,11 @@
 "nvidia-smi"
 
 ### 9
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"qwen3-4B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -53,7 +58,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-next-80B-A3B-8gpus.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-next-80B-A3B-8gpus.sh.txt
index fa31c876848..48497e0a2f3 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-next-80B-A3B-8gpus.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-next-80B-A3B-8gpus.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"qwen3-next-80B-A3B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,12 +60,12 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "awk"
 "{print $1}"
 "/root/mpi_rack_hostfile"
 
-### 11
+### 12
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-next-80B-A3B.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-next-80B-A3B.sh.txt
index 87689bcd220..d579cbfe6a5 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-next-80B-A3B.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-qwen3-next-80B-A3B.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"qwen3-next-80B-A3B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,12 +60,12 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "awk"
 "{print $1}"
 "/root/mpi_rack_hostfile"
 
-### 11
+### 12
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-qwen3.5-27B.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-qwen3.5-27B.sh.txt
index 53c86221e6a..58ee19fba5c 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-qwen3.5-27B.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-qwen3.5-27B.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"qwen3.5-27B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,7 +60,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-qwen3.5-35B-A3B-mtp.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-qwen3.5-35B-A3B-mtp.sh.txt
index 042f63869fa..e9e4f080a3d 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-qwen3.5-35B-A3B-mtp.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-qwen3.5-35B-A3B-mtp.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"qwen3.5-35B-A3B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,7 +60,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-qwen3.5-4B.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-qwen3.5-4B.sh.txt
index 49b84e687bb..4c4cb5787f2 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-qwen3.5-4B.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-qwen3.5-4B.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"qwen3.5-4B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,7 +60,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-qwen3.5-9B.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-qwen3.5-9B.sh.txt
index b44019d4a58..55e5bcccddf 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-qwen3.5-9B.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-qwen3.5-9B.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"qwen3.5-9B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,7 +60,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"
diff --git a/tests/snapshots/launch_scripts/sh/scripts/run-qwen3.6-27B.sh.txt b/tests/snapshots/launch_scripts/sh/scripts/run-qwen3.6-27B.sh.txt
index e3ec1d82697..a8a67450fd3 100644
--- a/tests/snapshots/launch_scripts/sh/scripts/run-qwen3.6-27B.sh.txt
+++ b/tests/snapshots/launch_scripts/sh/scripts/run-qwen3.6-27B.sh.txt
@@ -44,6 +44,11 @@
 "-m"
 
 ### 9
+"python3"
+"<REPO_ROOT>/scripts/../miles/utils/external_utils/model_args_utils.py"
+"qwen3.6-27B"
+
+### 10
 "ray"
 "start"
 "--head"
@@ -55,7 +60,7 @@
 "--dashboard-host=0.0.0.0"
 "--dashboard-port=8265"
 
-### 10
+### 11
 "ray"
 "job"
 "submit"