diff --git a/.github/workflows/benchmark-multinode-tmpl.yml b/.github/workflows/benchmark-multinode-tmpl.yml
index 6b5319c667..0a666a6e90 100644
--- a/.github/workflows/benchmark-multinode-tmpl.yml
+++ b/.github/workflows/benchmark-multinode-tmpl.yml
@@ -467,7 +467,7 @@ jobs:
         if: ${{ always() && (env.RUN_EVAL == 'true' || inputs.eval-only) }}
         uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
         with:
-          name: eval_${{ env.EXP_NAME }}_${{ env.RESULT_FILENAME }}
+          name: eval_${{ env.RESULT_FILENAME }}
           path: |
             meta_env.json
             results*.json
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-latency-dspark-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-latency-dspark-agentic.yaml
new file mode 100644
index 0000000000..7dc3d49793
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-latency-dspark-agentic.yaml
@@ -0,0 +1,111 @@
+name: "kimik3-vllm-agg-b200-tp8dp2-latency-dspark-agentic"
+
+# Two-node DEP profile: TP8 stays within each NVSwitch node, DP2 replicates
+# attention/dense across nodes, and EP16 shards the 896 experts across all GPUs.
+# This keeps FlashInfer fused collectives node-local while using shipped vLLM.
+# https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=b200&nodes=2&strategy=multi_node_dep
+model:
+  path: "kimik3"
+  container: "vllm/vllm-openai:nightly-b22afe45ac797ae58e67a7a3ad79ee5714024420@sha256:144356af876edbb3a4bfee23e1444b196cc3fdadd0a0c1a7f11f721756972a21"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "moonshotai/Kimi-K3"
+  container:
+    image: "vllm/vllm-openai:nightly-b22afe45ac797ae58e67a7a3ad79ee5714024420@sha256:144356af876edbb3a4bfee23e1444b196cc3fdadd0a0c1a7f11f721756972a21"
+
+dynamo:
+  install: false
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  interval_seconds: 10
+  max_attempts: 720
+
+resources:
+  gpu_type: "b200"
+  gpus_per_node: 8
+  agg_nodes: 2
+  agg_workers: 1
+  gpus_per_agg: 16
+
+infra:
+  etcd_nats_dedicated_node: false
+  nats_max_payload_mb: 32
+
+frontend:
+  type: vllm
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+  aggregated_environment:
+    # TP8 collectives stay within each B200 NVSwitch domain.
+    VLLM_ENABLE_K3_LATENT_MOE_TAIL_FUSION: "1"
+    VLLM_ALLREDUCE_USE_FLASHINFER: "1"
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800"
+    VLLM_USE_V2_MODEL_RUNNER: "1"
+    VLLM_USE_RUST_FRONTEND: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_LOG_STATS_INTERVAL: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_NET_DEVICES: "mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_10:1,mlx5_11:1"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    HUGGINGFACE_HUB_CACHE: "/hf_hub_cache"
+  vllm_config:
+    aggregated:
+      served-model-name: "moonshotai/Kimi-K3"
+      tensor-parallel-size: 8
+      data-parallel-size: 2
+      enable-expert-parallel: true
+      trust-remote-code: true
+      load-format: fastsafetensors
+      moe-backend: auto
+      # Keep headroom for the first FlashInfer MXFP4 MoE workspace.
+      gpu-memory-utilization: 0.90
+      no-enable-flashinfer-autotune: true
+      max-model-len: 1048576
+      kv-cache-dtype: fp8
+      attention-config: '{"use_prefill_query_quantization":true,"mla_prefill_backend":"flashinfer"}'
+      enable-prefix-caching: true
+      enable-prompt-tokens-details: true
+      enable-auto-tool-choice: true
+      tool-call-parser: kimi_k3
+      reasoning-parser: kimi_k3
+      language-model-only: true
+      max-num-seqs: 8
+      max-num-batched-tokens: 8192
+      # Throughput runs use the committed K7 golden synthetic AL 3.84. Eval
+      # selects the paired real-verification recipe instead.
+      speculative-config: '{"model":"Inferact/Kimi-K3-DSpark","num_speculative_tokens":7,"method":"dspark","attention_backend":"FLASHINFER_MLA","draft_sample_method":"probabilistic","rejection_sample_method":"synthetic","synthetic_acceptance_length":3.84}'
+
+sbatch_directives:
+  segment: "1"
+
+srun_options:
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  aiperf_server_metrics: true
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    AIPERF_TRACE_IDLE_GAP_CAP_SECONDS: "300"
+    AIPERF_LIVE_FAILED_REQUEST_THRESHOLD: "0.25"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-latency-dspark-eval-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-latency-dspark-eval-agentic.yaml
new file mode 100644
index 0000000000..f562682522
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-latency-dspark-eval-agentic.yaml
@@ -0,0 +1,109 @@
+name: "kimik3-vllm-agg-b200-tp8dp2-latency-dspark-eval-agentic"
+
+# Real-verification eval counterpart to the B200 DEP throughput recipe.
+# It keeps TP8/DP2/EP16 and target settings unchanged; only DSpark acceptance
+# changes from synthetic golden AL to target-verified block.
+model:
+  path: "kimik3"
+  container: "vllm/vllm-openai:nightly-b22afe45ac797ae58e67a7a3ad79ee5714024420@sha256:144356af876edbb3a4bfee23e1444b196cc3fdadd0a0c1a7f11f721756972a21"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "moonshotai/Kimi-K3"
+  container:
+    image: "vllm/vllm-openai:nightly-b22afe45ac797ae58e67a7a3ad79ee5714024420@sha256:144356af876edbb3a4bfee23e1444b196cc3fdadd0a0c1a7f11f721756972a21"
+
+dynamo:
+  install: false
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  interval_seconds: 10
+  max_attempts: 720
+
+resources:
+  gpu_type: "b200"
+  gpus_per_node: 8
+  agg_nodes: 2
+  agg_workers: 1
+  gpus_per_agg: 16
+
+infra:
+  etcd_nats_dedicated_node: false
+  nats_max_payload_mb: 32
+
+frontend:
+  type: vllm
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+  aggregated_environment:
+    # TP8 collectives stay within each B200 NVSwitch domain.
+    VLLM_ENABLE_K3_LATENT_MOE_TAIL_FUSION: "1"
+    VLLM_ALLREDUCE_USE_FLASHINFER: "1"
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800"
+    VLLM_USE_V2_MODEL_RUNNER: "1"
+    VLLM_USE_RUST_FRONTEND: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_LOG_STATS_INTERVAL: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_NET_DEVICES: "mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_10:1,mlx5_11:1"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    HUGGINGFACE_HUB_CACHE: "/hf_hub_cache"
+  vllm_config:
+    aggregated:
+      served-model-name: "moonshotai/Kimi-K3"
+      tensor-parallel-size: 8
+      data-parallel-size: 2
+      enable-expert-parallel: true
+      trust-remote-code: true
+      load-format: fastsafetensors
+      moe-backend: auto
+      # Keep headroom for the first FlashInfer MXFP4 MoE workspace.
+      gpu-memory-utilization: 0.90
+      no-enable-flashinfer-autotune: true
+      max-model-len: 1048576
+      kv-cache-dtype: fp8
+      attention-config: '{"use_prefill_query_quantization":true,"mla_prefill_backend":"flashinfer"}'
+      enable-prefix-caching: true
+      enable-prompt-tokens-details: true
+      enable-auto-tool-choice: true
+      tool-call-parser: kimi_k3
+      reasoning-parser: kimi_k3
+      language-model-only: true
+      max-num-seqs: 8
+      max-num-batched-tokens: 8192
+      # Eval must verify every draft against real target logits.
+      speculative-config: '{"model":"Inferact/Kimi-K3-DSpark","num_speculative_tokens":7,"method":"dspark","attention_backend":"FLASHINFER_MLA","draft_sample_method":"probabilistic","rejection_sample_method":"block"}'
+
+sbatch_directives:
+  segment: "1"
+
+srun_options:
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  aiperf_server_metrics: true
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    AIPERF_TRACE_IDLE_GAP_CAP_SECONDS: "300"
+    AIPERF_LIVE_FAILED_REQUEST_THRESHOLD: "0.25"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-vllm-simple-offload-dspark-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-vllm-simple-offload-dspark-agentic.yaml
new file mode 100644
index 0000000000..0119d79219
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-vllm-simple-offload-dspark-agentic.yaml
@@ -0,0 +1,112 @@
+name: "kimik3-vllm-agg-b200-tp8dp2-simple-offload-dspark-agentic"
+
+# Two-node DEP profile: TP8 stays within each NVSwitch node, DP2 replicates
+# attention/dense across nodes, and EP16 shards the 896 experts across all GPUs.
+# This keeps FlashInfer fused collectives node-local while using shipped vLLM.
+# https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=b200&nodes=2&strategy=multi_node_dep
+model:
+  path: "kimik3"
+  container: "vllm/vllm-openai:nightly-b22afe45ac797ae58e67a7a3ad79ee5714024420@sha256:144356af876edbb3a4bfee23e1444b196cc3fdadd0a0c1a7f11f721756972a21"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "moonshotai/Kimi-K3"
+  container:
+    image: "vllm/vllm-openai:nightly-b22afe45ac797ae58e67a7a3ad79ee5714024420@sha256:144356af876edbb3a4bfee23e1444b196cc3fdadd0a0c1a7f11f721756972a21"
+
+dynamo:
+  install: false
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  interval_seconds: 10
+  max_attempts: 720
+
+resources:
+  gpu_type: "b200"
+  gpus_per_node: 8
+  agg_nodes: 2
+  agg_workers: 1
+  gpus_per_agg: 16
+
+infra:
+  etcd_nats_dedicated_node: false
+  nats_max_payload_mb: 32
+
+frontend:
+  type: vllm
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+  aggregated_environment:
+    # TP8 collectives stay within each B200 NVSwitch domain.
+    VLLM_ENABLE_K3_LATENT_MOE_TAIL_FUSION: "1"
+    VLLM_ALLREDUCE_USE_FLASHINFER: "1"
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800"
+    VLLM_USE_V2_MODEL_RUNNER: "1"
+    VLLM_USE_RUST_FRONTEND: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_LOG_STATS_INTERVAL: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_NET_DEVICES: "mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_10:1,mlx5_11:1"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    HUGGINGFACE_HUB_CACHE: "/hf_hub_cache"
+  vllm_config:
+    aggregated:
+      served-model-name: "moonshotai/Kimi-K3"
+      tensor-parallel-size: 8
+      data-parallel-size: 2
+      enable-expert-parallel: true
+      trust-remote-code: true
+      load-format: fastsafetensors
+      moe-backend: auto
+      # Keep headroom for the first FlashInfer MXFP4 MoE workspace.
+      gpu-memory-utilization: 0.90
+      no-enable-flashinfer-autotune: true
+      max-model-len: 1048576
+      kv-cache-dtype: fp8
+      attention-config: '{"use_prefill_query_quantization":true,"mla_prefill_backend":"flashinfer"}'
+      enable-prefix-caching: true
+      enable-prompt-tokens-details: true
+      enable-auto-tool-choice: true
+      tool-call-parser: kimi_k3
+      reasoning-parser: kimi_k3
+      language-model-only: true
+      max-num-seqs: 8
+      max-num-batched-tokens: 8192
+      # Throughput runs use the committed K7 golden synthetic AL 3.84. Eval
+      # selects the paired real-verification recipe instead.
+      speculative-config: '{"model":"Inferact/Kimi-K3-DSpark","num_speculative_tokens":7,"method":"dspark","attention_backend":"FLASHINFER_MLA","draft_sample_method":"probabilistic","rejection_sample_method":"synthetic","synthetic_acceptance_length":3.84}'
+      kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use_per_rank":236223201280,"lazy_offload":false}}'
+
+sbatch_directives:
+  segment: "1"
+
+srun_options:
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  aiperf_server_metrics: true
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    AIPERF_TRACE_IDLE_GAP_CAP_SECONDS: "300"
+    AIPERF_LIVE_FAILED_REQUEST_THRESHOLD: "0.25"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-vllm-simple-offload-dspark-eval-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-vllm-simple-offload-dspark-eval-agentic.yaml
new file mode 100644
index 0000000000..a3ae8342b9
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-vllm-simple-offload-dspark-eval-agentic.yaml
@@ -0,0 +1,110 @@
+name: "kimik3-vllm-agg-b200-tp8dp2-simple-offload-dspark-eval-agentic"
+
+# Real-verification eval counterpart to the B200 DEP offload throughput recipe.
+# It keeps TP8/DP2/EP16 and the CPU KV connector unchanged; only DSpark
+# acceptance changes from synthetic golden AL to target-verified block.
+model:
+  path: "kimik3"
+  container: "vllm/vllm-openai:nightly-b22afe45ac797ae58e67a7a3ad79ee5714024420@sha256:144356af876edbb3a4bfee23e1444b196cc3fdadd0a0c1a7f11f721756972a21"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "moonshotai/Kimi-K3"
+  container:
+    image: "vllm/vllm-openai:nightly-b22afe45ac797ae58e67a7a3ad79ee5714024420@sha256:144356af876edbb3a4bfee23e1444b196cc3fdadd0a0c1a7f11f721756972a21"
+
+dynamo:
+  install: false
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  interval_seconds: 10
+  max_attempts: 720
+
+resources:
+  gpu_type: "b200"
+  gpus_per_node: 8
+  agg_nodes: 2
+  agg_workers: 1
+  gpus_per_agg: 16
+
+infra:
+  etcd_nats_dedicated_node: false
+  nats_max_payload_mb: 32
+
+frontend:
+  type: vllm
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+  aggregated_environment:
+    # TP8 collectives stay within each B200 NVSwitch domain.
+    VLLM_ENABLE_K3_LATENT_MOE_TAIL_FUSION: "1"
+    VLLM_ALLREDUCE_USE_FLASHINFER: "1"
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800"
+    VLLM_USE_V2_MODEL_RUNNER: "1"
+    VLLM_USE_RUST_FRONTEND: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_LOG_STATS_INTERVAL: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_NET_DEVICES: "mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_10:1,mlx5_11:1"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    HUGGINGFACE_HUB_CACHE: "/hf_hub_cache"
+  vllm_config:
+    aggregated:
+      served-model-name: "moonshotai/Kimi-K3"
+      tensor-parallel-size: 8
+      data-parallel-size: 2
+      enable-expert-parallel: true
+      trust-remote-code: true
+      load-format: fastsafetensors
+      moe-backend: auto
+      # Keep headroom for the first FlashInfer MXFP4 MoE workspace.
+      gpu-memory-utilization: 0.90
+      no-enable-flashinfer-autotune: true
+      max-model-len: 1048576
+      kv-cache-dtype: fp8
+      attention-config: '{"use_prefill_query_quantization":true,"mla_prefill_backend":"flashinfer"}'
+      enable-prefix-caching: true
+      enable-prompt-tokens-details: true
+      enable-auto-tool-choice: true
+      tool-call-parser: kimi_k3
+      reasoning-parser: kimi_k3
+      language-model-only: true
+      max-num-seqs: 8
+      max-num-batched-tokens: 8192
+      # Eval must verify every draft against real target logits.
+      speculative-config: '{"model":"Inferact/Kimi-K3-DSpark","num_speculative_tokens":7,"method":"dspark","attention_backend":"FLASHINFER_MLA","draft_sample_method":"probabilistic","rejection_sample_method":"block"}'
+      kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use_per_rank":236223201280,"lazy_offload":false}}'
+
+sbatch_directives:
+  segment: "1"
+
+srun_options:
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  aiperf_server_metrics: true
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    AIPERF_TRACE_IDLE_GAP_CAP_SECONDS: "300"
+    AIPERF_LIVE_FAILED_REQUEST_THRESHOLD: "0.25"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index ab43c7573a..c752b04f67 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8035,6 +8035,90 @@ kimik3-fp4-b200-dynamo-vllm-agentic:
           ep: 1
           dp-attn: false
 
+# DSpark uses two node-local TP8 attention/dense groups and DP2 across nodes.
+# Expert parallel spans all 16 ranks, so the 896 experts remain EP16 while
+# FlashInfer fused collectives stay inside each B200 NVSwitch domain.
+kimik3-fp4-b200-dynamo-vllm-agentic-dspark:
+  image: vllm/vllm-openai:nightly-b22afe45ac797ae58e67a7a3ad79ee5714024420@sha256:144356af876edbb3a4bfee23e1444b196cc3fdadd0a0c1a7f11f721756972a21
+  model: moonshotai/Kimi-K3
+  model-prefix: kimik3
+  runner: cluster:b200-dgxc
+  precision: fp4
+  framework: dynamo-vllm
+  multinode: true
+  disagg: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.63
+      search-space:
+      # Low-latency and high-interactivity points.
+      - spec-decoding: mtp
+        conc-list: [1, 2, 4]
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 16
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-latency-dspark-agentic.yaml"
+          - "EVAL_CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-latency-dspark-eval-agentic.yaml"
+        decode:
+          num-worker: 0
+          tp: 8
+          ep: 16
+          dp-attn: true
+      # Balanced medium-concurrency points.
+      - spec-decoding: mtp
+        conc-list: [8, 12, 16]
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 16
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-latency-dspark-agentic.yaml"
+          - "EVAL_CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-latency-dspark-eval-agentic.yaml"
+        decode:
+          num-worker: 0
+          tp: 8
+          ep: 16
+          dp-attn: true
+      # GPU-resident throughput points around the prior c16-c32 KV cliff.
+      - spec-decoding: mtp
+        conc-list: [20, 24, 28, 32]
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 16
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-latency-dspark-agentic.yaml"
+          - "EVAL_CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-latency-dspark-eval-agentic.yaml"
+        decode:
+          num-worker: 0
+          tp: 8
+          ep: 16
+          dp-attn: true
+      # CPU KV-offload crossover and capacity points. Keep the resident points
+      # above so the same concurrency can be compared with one variable changed.
+      - spec-decoding: mtp
+        kv-offloading: dram
+        kv-offload-backend: { name: vllm-simple, version: "13c59a3" }
+        conc-list: [8, 12, 16, 20, 24, 28, 32, 48, 64]
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 16
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-vllm-simple-offload-dspark-agentic.yaml"
+          - "EVAL_CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-vllm-simple-offload-dspark-eval-agentic.yaml"
+        decode:
+          num-worker: 0
+          tp: 8
+          ep: 16
+          dp-attn: true
+
 qwen3.5-fp8-gb300-dynamo-sglang-mtp:
   image: lmsysorg/sglang:v0.5.14-cu130@sha256:5027e95bf6ec536856b1b52a91d1f35ff5c564ab83e8a94758a169ff09bb8df3
   model: Qwen/Qwen3.5-397B-A17B-FP8
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 23369f7273..0bce8d894f 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5784,7 +5784,6 @@
     - "Update the B300 DSV4 AgentX MTP image and SimpleCPU sweep through c576."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2536
   
-
 - config-keys:
     - kimik3-fp4-b300-vllm-agentic-dspark
   description:
@@ -5803,4 +5802,19 @@
     - "Use Dynamo header-based session affinity and AIPerf agentx-v1.0.1 (b7b16cf851885567988a643282266bce74e34437)."
     - "Enable SGLang engine metrics and require the sglang: server-metric prefix so the published trace artifacts carry backend data."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2364
-  
+
+- config-keys:
+    - kimik3-fp4-b200-dynamo-vllm-agentic-dspark
+  description:
+    - "Add Kimi K3 B200 AgentX latency, balanced, and throughput tiers on a two-node TP8/DP2/EP16 DSpark topology with tensor parallelism local to each B200"
+    - "Add a TP8/DP2/EP16 SimpleCPUOffloadConnector capacity tier with 220 GiB of host KV per rank"
+    - "Keep throughput on DSpark level 7 probabilistic drafting with committed golden synthetic AL 3.84; EVAL_ONLY selects committed rejection_sample_method=block recipes for real target verification"
+    - "Use the digest-pinned 2026-08-10 upstream vLLM nightly and direct Inferact/Kimi-K3-DSpark model ID as shipped, with no local compatibility layer or installed serving-engine rewrite"
+    - "Use node-local TP8 FlashInfer collectives and cross-node DP2/EP16 so the digest-pinned upstream serving engine runs without a compatibility layer or source rewrite"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2475
+
+- config-keys:
+    - kimik3-fp4-b200-dynamo-vllm-agentic-dspark
+  description:
+    - "Extend the SimpleCPUOffloadConnector grid to c8/c12/c16/c20/c24/c28/c32/c48/c64 to locate its crossover against the resident curve"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2475
diff --git a/runners/launch_b200-dgxc.sh b/runners/launch_b200-dgxc.sh
index 2cd2f2ee81..8c2b357b00 100644
--- a/runners/launch_b200-dgxc.sh
+++ b/runners/launch_b200-dgxc.sh
@@ -165,21 +165,16 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
         rm -rf "$SRT_REPO_DIR"
     fi
 
-    # TODO(CJQ): make first class upon srt-slurm upstream refactor
+    # Kimi K3 aggregate profiles use the srt-slurm fork that supports direct
+    # multi-node vLLM. Pin the tested renderer so branch movement cannot change
+    # generated rank commands between sweep points.
     if [[ "$IS_AGENTIC" == "1" && $MODEL_PREFIX == "kimik3" ]]; then
-        # Direct-vLLM agentic experiment (Variant D): srt-slurm PR #278
-        # (kylliang/direct-aggregate-vllm) adds frontend.type: vllm — `vllm
-        # serve` owns the OpenAI port itself, no Dynamo layer. The fork branch
-        # carries PR #278 plus the multi-node extension (vLLM-native
-        # --master-addr/--nnodes/--node-rank serve + headless non-leader
-        # ranks) so the 2-node TP8xPP2 topology can run.
         git clone --branch klaud/direct-vllm-multinode --single-branch https://github.com/functionstackx/srt-slurm-nv.git "$SRT_REPO_DIR" || exit 1
         cd "$SRT_REPO_DIR" || exit 1
-        if [[ $MODEL_PREFIX == "kimik3" ]]; then
-            mkdir -p recipes/vllm/kimi-k3/agentic || exit 1
-            cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic" \
-                recipes/vllm/kimi-k3/agentic || exit 1
-        fi
+        git checkout df5baa93f4caf5169dea2a4236ad2cc742fe40e7 || exit 1
+        mkdir -p recipes/vllm/kimi-k3/agentic || exit 1
+        cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic" \
+            recipes/vllm/kimi-k3/agentic || exit 1
     elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "dsv4" ]]; then
         git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
         cd "$SRT_REPO_DIR" || exit 1
@@ -352,6 +347,15 @@ EOF
     echo "MODEL_PATH=$MODEL_PATH (exists=$(test -d "$MODEL_PATH" && echo yes || echo NO))"
     ls -ld "$MODEL_PATH" 2>&1 || ls /lustre/fsw/models/ 2>&1 | head -40
 
+    # An eval row may point at a committed real-verification recipe while its
+    # throughput row keeps synthetic golden acceptance. Only configs that set
+    # EVAL_CONFIG_FILE opt into this selection; all other configs keep using
+    # CONFIG_FILE unchanged.
+    if [[ "${EVAL_ONLY:-false}" == "true" && -n "${EVAL_CONFIG_FILE:-}" ]]; then
+        CONFIG_FILE="$EVAL_CONFIG_FILE"
+        echo "EVAL_ONLY=true: selecting real-verification recipe $CONFIG_FILE"
+    fi
+
     if [[ -z "$CONFIG_FILE" ]]; then
         echo "Error: CONFIG_FILE is not set. The srt-slurm path requires a CONFIG_FILE in additional-settings." >&2
         echo "Config: MODEL_PREFIX=${MODEL_PREFIX} PRECISION=${PRECISION} FRAMEWORK=${FRAMEWORK}" >&2