diff --git a/docs/models/qwen/qwen35-vl.md b/docs/models/qwen/qwen35-vl.md
index 7b3cee0dea..999640b9f3 100644
--- a/docs/models/qwen/qwen35-vl.md
+++ b/docs/models/qwen/qwen35-vl.md
@@ -52,6 +52,29 @@ Please upgrade to `transformers` >= 5.2.0 in order to use the Qwen 3.5 models.
 
 For checkpoint conversion, inference, finetuning recipes, and step-by-step training guides, see the [Qwen 3.5 Examples](https://github.com/NVIDIA-NeMo/Megatron-Bridge/blob/main/examples/models/qwen/qwen35_vl/README.md).
 
+### Text-only pretraining
+
+The Qwen3.5 9B and 35B-A3B recipes can pretrain only the language-model
+component of the unified models. They derive the registered language-model
+provider from the nested Hugging Face `text_config`; no vision model,
+projection, processor, or multimodal dataset is created.
+
+```python
+from megatron.bridge.recipes.qwen import qwen35_text_9b_pretrain_config
+
+config = qwen35_text_9b_pretrain_config()
+```
+
+The canonical aliases select eight-GPU GB200 BF16 library recipes. The dense
+9B recipe, `qwen35_text_9b_pretrain_8gpu_gb200_bf16_config`, uses the same data
+parallel topology as the Llama 3 8B GB200 performance recipe, with
+module-scoped CUDA graphs so library correctness checks remain enabled. The MoE recipe,
+`qwen35_text_35b_a3b_pretrain_8gpu_gb200_bf16_config`, uses the applicable Qwen3.5-VL
+GB200 HybridEP settings with learned routing. Both retain library-recipe
+training, evaluation, logging, checkpointing, and correctness defaults. Set
+`config.dataset.blend` (or `config.dataset.data_path`) to use a prepared
+Megatron indexed text dataset.
+
 ## Hugging Face Model Cards
 
 - Qwen3.5 0.8B: https://huggingface.co/Qwen/Qwen3.5-0.8B
diff --git a/src/megatron/bridge/recipes/nemotronh/__init__.py b/src/megatron/bridge/recipes/nemotronh/__init__.py
index 4a3ab0a63c..ed1375a11a 100644
--- a/src/megatron/bridge/recipes/nemotronh/__init__.py
+++ b/src/megatron/bridge/recipes/nemotronh/__init__.py
@@ -14,6 +14,10 @@
 
 # Nemotron Nano v2 models
 # Nemotron 3 Nano models
+from megatron.bridge.recipes.nemotronh.gb200 import (
+    nemotron_3_nano_gb200_pretrain_config,
+    nemotron_3_nano_pretrain_8gpu_gb200_bf16_config,
+)
 from megatron.bridge.recipes.nemotronh.nemotron_3_nano import (
     nemotron_3_nano_peft_config,
     nemotron_3_nano_pretrain_config,
@@ -88,6 +92,8 @@
     "nemotron_3_nano_pretrain_config",
     "nemotron_3_nano_sft_config",
     "nemotron_3_nano_peft_config",
+    "nemotron_3_nano_gb200_pretrain_config",
+    "nemotron_3_nano_pretrain_8gpu_gb200_bf16_config",
     # Nemotron 3 Nano 4B model
     "nemotron_3_nano_4b_pretrain_config",
     "nemotron_3_nano_4b_sft_config",
diff --git a/src/megatron/bridge/recipes/nemotronh/gb200/__init__.py b/src/megatron/bridge/recipes/nemotronh/gb200/__init__.py
new file mode 100644
index 0000000000..c4b08f394a
--- /dev/null
+++ b/src/megatron/bridge/recipes/nemotronh/gb200/__init__.py
@@ -0,0 +1,24 @@
+# Copyright (c) 2026, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from megatron.bridge.recipes.nemotronh.gb200.nemotron_3_nano import (
+    nemotron_3_nano_gb200_pretrain_config,
+    nemotron_3_nano_pretrain_8gpu_gb200_bf16_config,
+)
+
+
+__all__ = [
+    "nemotron_3_nano_gb200_pretrain_config",
+    "nemotron_3_nano_pretrain_8gpu_gb200_bf16_config",
+]
diff --git a/src/megatron/bridge/recipes/nemotronh/gb200/nemotron_3_nano.py b/src/megatron/bridge/recipes/nemotronh/gb200/nemotron_3_nano.py
new file mode 100644
index 0000000000..fcb5070d76
--- /dev/null
+++ b/src/megatron/bridge/recipes/nemotronh/gb200/nemotron_3_nano.py
@@ -0,0 +1,148 @@
+# Copyright (c) 2026, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""GB200 pretraining recipe for Nemotron 3 Nano."""
+
+import torch
+
+from megatron.bridge import AutoBridge
+from megatron.bridge.recipes.common import _pretrain_common
+from megatron.bridge.training.comm_overlap import CommOverlapConfig
+from megatron.bridge.training.config import ConfigContainer
+from megatron.bridge.training.mixed_precision import get_mixed_precision_config
+from megatron.bridge.utils.cuda_graph import set_cuda_graph_modules
+
+
+_NEMOTRON_3_NANO_MODEL_ID = "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16"
+
+
+def nemotron_3_nano_pretrain_8gpu_gb200_bf16_config() -> ConfigContainer:
+    """Return the Nemotron 3 Nano BF16 pretraining config for eight GB200 GPUs.
+
+    The recipe retains the established optimizer, scheduler, routing, and BF16
+    contracts. It applies the validated GB200 TP1/EP8 HybridEP topology and
+    uses a 4,096-token sequence length for the paired NeMo-CI convergence
+    workload.
+
+    Returns:
+        GB200 BF16 pretraining configuration.
+    """
+    cfg = _pretrain_common()
+
+    cfg.model = AutoBridge.from_hf_pretrained(_NEMOTRON_3_NANO_MODEL_ID).to_megatron_provider(load_weights=False)
+    # Pretraining may use a tokenizer other than the HF checkpoint tokenizer.
+    # Defer the model vocabulary size to the runtime tokenizer, matching the
+    # pre-migration MambaModelProvider recipe behavior.
+    cfg.model.vocab_size = None
+    cfg.tokenizer.tokenizer_model = _NEMOTRON_3_NANO_MODEL_ID
+
+    cfg.model.seq_length = 4096
+    cfg.dataset.seq_length = 4096
+    cfg.dataset.blend = None
+    cfg.dataset.num_workers = 8
+    cfg.dataset.mmap_bin_files = False
+
+    cfg.model.tensor_model_parallel_size = 1
+    cfg.model.pipeline_model_parallel_size = 1
+    cfg.model.pipeline_model_parallel_layout = None
+    cfg.model.pipeline_dtype = torch.bfloat16
+    cfg.model.virtual_pipeline_model_parallel_size = None
+    cfg.model.context_parallel_size = 1
+    cfg.model.sequence_parallel = False
+    cfg.model.expert_tensor_parallel_size = 1
+    cfg.model.expert_model_parallel_size = 8
+
+    cfg.model.moe_token_dispatcher_type = "flex"
+    cfg.model.moe_flex_dispatcher_backend = "hybridep"
+    cfg.model.moe_flex_dispatcher_num_sms = 16
+    cfg.model.moe_hybridep_num_sms = None
+    cfg.model.moe_shared_expert_overlap = False
+    cfg.model.moe_router_force_load_balancing = False
+
+    cfg.train.train_iters = 39735
+    cfg.train.global_batch_size = 3072
+    cfg.train.micro_batch_size = 2
+    cfg.train.manual_gc = False
+    cfg.train.manual_gc_interval = 0
+
+    cfg.model.transformer_impl = "transformer_engine"
+
+    # Match the validated GB200 performance recipe's TE-scoped graph set.
+    cfg.model.cuda_graph_impl = "transformer_engine"
+    set_cuda_graph_modules(cfg.model, ["attn", "mamba", "moe_router", "moe_preprocess"])
+    cfg.model.cuda_graph_warmup_steps = 3
+    cfg.model.use_te_rng_tracker = True
+    cfg.rng.te_rng_tracker = True
+
+    cfg.model.attention_backend = "fused"
+    cfg.model.moe_router_fusion = False
+    cfg.model.moe_permute_fusion = True
+    cfg.model.moe_grouped_gemm = True
+    cfg.model.cross_entropy_loss_fusion = True
+    cfg.model.apply_rope_fusion = True
+    cfg.model.cross_entropy_fusion_impl = "native"
+    cfg.model.recompute_granularity = None
+    cfg.model.recompute_modules = None
+    cfg.model.fine_grained_activation_offloading = False
+    cfg.model.offload_modules = None
+    cfg.model.moe_router_padding_for_fp8 = False
+    cfg.rerun_state_machine.check_for_nan_in_loss = False
+
+    cfg.optimizer.use_precision_aware_optimizer = False
+    cfg.optimizer.main_grads_dtype = torch.float32
+    cfg.optimizer.main_params_dtype = torch.float32
+    cfg.optimizer.exp_avg_dtype = torch.float32
+    cfg.optimizer.exp_avg_sq_dtype = torch.float32
+    cfg.optimizer.lr = 1.6e-3
+    cfg.optimizer.weight_decay = 0.1
+    cfg.optimizer.min_lr = 1.6e-5
+    cfg.scheduler.lr_warmup_iters = 333
+
+    # Keep BF16 compute while reducing gradients in BF16 instead of FP32.
+    cfg.mixed_precision = get_mixed_precision_config(cfg.mixed_precision)
+    cfg.mixed_precision.grad_reduce_in_fp32 = False
+
+    cfg.comm_overlap = CommOverlapConfig(
+        tp_comm_bootstrap_backend="nccl",
+        tp_comm_overlap=False,
+    )
+    cfg.comm_overlap.delay_wgrad_compute = False
+    cfg.comm_overlap.overlap_moe_expert_parallel_comm = False
+
+    cfg.checkpoint.save_interval = 200
+    cfg.checkpoint.ckpt_assume_constant_structure = True
+    cfg.checkpoint.dist_ckpt_strictness = "log_all"
+
+    cfg.ddp.overlap_grad_reduce = True
+    cfg.ddp.overlap_param_gather = True
+    cfg.ddp.check_for_nan_in_grad = False
+    cfg.ddp.use_distributed_optimizer = True
+    cfg.ddp.grad_reduce_in_fp32 = False
+
+    cfg.model.init_method_std = 0.0173
+    cfg.model.use_fused_weighted_squared_relu = True
+
+    return cfg
+
+
+# NeMo-CI appends ``_pretrain_config`` to MODEL_RECIPE_NAME. This explicit
+# alias lets the GB200 release case select the hardware recipe without changing
+# the legacy ``nemotron_3_nano_pretrain_config`` default.
+nemotron_3_nano_gb200_pretrain_config = nemotron_3_nano_pretrain_8gpu_gb200_bf16_config
+
+
+__all__ = [
+    "nemotron_3_nano_gb200_pretrain_config",
+    "nemotron_3_nano_pretrain_8gpu_gb200_bf16_config",
+]
diff --git a/src/megatron/bridge/recipes/qwen/__init__.py b/src/megatron/bridge/recipes/qwen/__init__.py
index b228b95ab4..93588f4e14 100644
--- a/src/megatron/bridge/recipes/qwen/__init__.py
+++ b/src/megatron/bridge/recipes/qwen/__init__.py
@@ -12,6 +12,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+# Qwen3.5 GB200 models
+from .gb200.qwen35 import (
+    qwen35_text_9b_pretrain_8gpu_gb200_bf16_config,
+    qwen35_text_35b_a3b_pretrain_8gpu_gb200_bf16_config,
+)
+
 # Qwen2 models
 from .qwen2 import (
     qwen2_1p5b_peft_config,
@@ -88,6 +94,9 @@
     qwen3_next_80b_a3b_sft_config,
 )
 
+# Qwen3.5 text models
+from .qwen35 import qwen35_text_9b_pretrain_config, qwen35_text_35b_a3b_pretrain_config
+
 
 __all__ = [
     # Qwen2 models
@@ -155,4 +164,9 @@
     "qwen3_next_80b_a3b_pretrain_config",
     "qwen3_next_80b_a3b_sft_config",
     "qwen3_next_80b_a3b_peft_config",
+    # Qwen3.5 text models
+    "qwen35_text_9b_pretrain_config",
+    "qwen35_text_9b_pretrain_8gpu_gb200_bf16_config",
+    "qwen35_text_35b_a3b_pretrain_config",
+    "qwen35_text_35b_a3b_pretrain_8gpu_gb200_bf16_config",
 ]
diff --git a/src/megatron/bridge/recipes/qwen/gb200/__init__.py b/src/megatron/bridge/recipes/qwen/gb200/__init__.py
new file mode 100644
index 0000000000..2d5eeb765f
--- /dev/null
+++ b/src/megatron/bridge/recipes/qwen/gb200/__init__.py
@@ -0,0 +1,24 @@
+# Copyright (c) 2026, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from megatron.bridge.recipes.qwen.gb200.qwen35 import (
+    qwen35_text_9b_pretrain_8gpu_gb200_bf16_config,
+    qwen35_text_35b_a3b_pretrain_8gpu_gb200_bf16_config,
+)
+
+
+__all__ = [
+    "qwen35_text_9b_pretrain_8gpu_gb200_bf16_config",
+    "qwen35_text_35b_a3b_pretrain_8gpu_gb200_bf16_config",
+]
diff --git a/src/megatron/bridge/recipes/qwen/gb200/qwen35.py b/src/megatron/bridge/recipes/qwen/gb200/qwen35.py
new file mode 100644
index 0000000000..f151a90b01
--- /dev/null
+++ b/src/megatron/bridge/recipes/qwen/gb200/qwen35.py
@@ -0,0 +1,191 @@
+# Copyright (c) 2026, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""GB200 text-only pretraining recipes for Qwen3.5 dense and MoE models."""
+
+from __future__ import annotations
+
+import torch
+from transformers import AutoConfig
+
+from megatron.bridge import AutoBridge
+from megatron.bridge.recipes.common import _pretrain_common
+from megatron.bridge.training.comm_overlap import CommOverlapConfig
+from megatron.bridge.training.config import ConfigContainer
+from megatron.bridge.training.mixed_precision import bf16_mixed
+
+
+_QWEN35_9B_BASE = "Qwen/Qwen3.5-9B-Base"
+_QWEN35_35B_A3B_BASE = "Qwen/Qwen3.5-35B-A3B-Base"
+
+
+def qwen35_text_9b_pretrain_8gpu_gb200_bf16_config() -> ConfigContainer:
+    """Return a text-only Qwen3.5-9B pretraining config for eight GB200 GPUs."""
+    cfg = _pretrain_common()
+
+    text_config = AutoConfig.from_pretrained(_QWEN35_9B_BASE).text_config
+    # The nested text config intentionally omits ``architectures``. AutoBridge
+    # needs it to select the registered causal-LM bridge instead of the VLM.
+    text_config.architectures = ["Qwen3_5ForCausalLM"]
+    cfg.model = AutoBridge.from_hf_config(text_config).to_megatron_provider(load_weights=False)
+    cfg.tokenizer.tokenizer_model = _QWEN35_9B_BASE
+    cfg.dataset.seq_length = 4096
+    cfg.dataset.blend = None
+    cfg.dataset.num_workers = 8
+
+    # Follow the Llama 3 8B GB200 topology: keep model parallelism at one and
+    # use all eight GPUs for data parallelism.
+    cfg.model.tensor_model_parallel_size = 1
+    cfg.model.pipeline_model_parallel_size = 1
+    cfg.model.pipeline_model_parallel_layout = None
+    cfg.model.pipeline_dtype = torch.bfloat16
+    cfg.model.virtual_pipeline_model_parallel_size = None
+    cfg.model.context_parallel_size = 1
+    cfg.model.expert_model_parallel_size = 1
+    cfg.model.expert_tensor_parallel_size = 1
+    cfg.model.sequence_parallel = False
+    cfg.model.seq_length = 4096
+    cfg.model.init_method_std = 0.02
+    cfg.train.global_batch_size = 128
+    cfg.train.micro_batch_size = 2
+
+    cfg.model.transformer_impl = "transformer_engine"
+    cfg.model.bias_activation_fusion = True
+    cfg.model.cross_entropy_loss_fusion = True
+    cfg.model.cross_entropy_fusion_impl = "native"
+    cfg.model.apply_rope_fusion = True
+
+    cfg.model.recompute_granularity = None
+    cfg.model.recompute_method = None
+    cfg.model.recompute_num_layers = None
+    cfg.model.recompute_modules = None
+    cfg.model.fine_grained_activation_offloading = False
+    cfg.model.offload_modules = None
+
+    # Capture the dense attention and MLP modules. Keep cross entropy on the
+    # native fused path validated by the 64-GPU GB200 performance run.
+    cfg.model.cuda_graph_impl = "transformer_engine"
+    cfg.model.cuda_graph_scope = None
+    cfg.model.cuda_graph_modules = ["attn", "mlp"]
+    cfg.model.cuda_graph_warmup_steps = 3
+    cfg.model.use_te_rng_tracker = True
+    cfg.rng.te_rng_tracker = True
+
+    cfg.optimizer.use_precision_aware_optimizer = False
+    cfg.optimizer.main_grads_dtype = torch.float32
+    cfg.optimizer.main_params_dtype = torch.float32
+    cfg.optimizer.exp_avg_dtype = torch.float32
+    cfg.optimizer.exp_avg_sq_dtype = torch.float32
+
+    cfg.mixed_precision = bf16_mixed()
+    cfg.mixed_precision.grad_reduce_in_fp32 = False
+
+    cfg.ddp.overlap_grad_reduce = True
+    cfg.ddp.overlap_param_gather = True
+    cfg.ddp.grad_reduce_in_fp32 = False
+    cfg.ddp.check_for_nan_in_grad = False
+    cfg.ddp.use_distributed_optimizer = True
+    cfg.ddp.use_megatron_fsdp = False
+    cfg.rerun_state_machine.check_for_nan_in_loss = False
+
+    cfg.comm_overlap = CommOverlapConfig(tp_comm_overlap=False)
+    return cfg
+
+
+def qwen35_text_35b_a3b_pretrain_8gpu_gb200_bf16_config() -> ConfigContainer:
+    """Return a text-only Qwen3.5-35B-A3B pretraining config for eight GB200 GPUs."""
+    cfg = _pretrain_common()
+
+    text_config = AutoConfig.from_pretrained(_QWEN35_35B_A3B_BASE).text_config
+    # The nested text config intentionally omits ``architectures``. AutoBridge
+    # needs it to select the registered causal-LM bridge instead of the VLM.
+    text_config.architectures = ["Qwen3_5MoeForCausalLM"]
+    cfg.model = AutoBridge.from_hf_config(text_config).to_megatron_provider(load_weights=False)
+    cfg.tokenizer.tokenizer_model = _QWEN35_35B_A3B_BASE
+    cfg.dataset.seq_length = 4096
+    cfg.dataset.blend = None
+    cfg.dataset.num_workers = 8
+
+    # Match the Qwen3.5-VL GB200 topology while training only the text model.
+    cfg.model.tensor_model_parallel_size = 1
+    cfg.model.pipeline_model_parallel_size = 1
+    cfg.model.pipeline_model_parallel_layout = None
+    cfg.model.pipeline_dtype = torch.bfloat16
+    cfg.model.virtual_pipeline_model_parallel_size = None
+    cfg.model.context_parallel_size = 1
+    cfg.model.expert_model_parallel_size = 8
+    cfg.model.expert_tensor_parallel_size = 1
+    cfg.model.sequence_parallel = False
+    cfg.model.seq_length = 4096
+    cfg.model.init_method_std = 0.02
+    cfg.train.global_batch_size = 512
+    # MBS4 is suitable for force-balanced throughput benchmarking, but OOMs
+    # with learned routing. MBS1 was validated with real RP2 data on GB200.
+    cfg.train.micro_batch_size = 1
+
+    cfg.model.transformer_impl = "transformer_engine"
+    cfg.model.bias_activation_fusion = True
+    cfg.model.moe_router_fusion = True
+    cfg.model.moe_permute_fusion = True
+    cfg.model.moe_grouped_gemm = True
+    cfg.model.cross_entropy_loss_fusion = True
+    # Keep the library-safe native implementation instead of the performance
+    # harness's TE cross-entropy path, which currently warns about stability.
+    cfg.model.cross_entropy_fusion_impl = "native"
+    cfg.model.apply_rope_fusion = True
+
+    cfg.model.recompute_granularity = None
+    cfg.model.recompute_method = None
+    cfg.model.recompute_num_layers = None
+    cfg.model.recompute_modules = None
+    cfg.model.fine_grained_activation_offloading = False
+    cfg.model.offload_modules = None
+
+    # Fixed-length text batches can use the scopes that the VLM recipe must
+    # disable for variable-length multimodal inputs.
+    cfg.model.cuda_graph_impl = "transformer_engine"
+    cfg.model.cuda_graph_scope = None
+    cfg.model.cuda_graph_modules = ["attn", "moe_router", "moe_preprocess"]
+    cfg.model.cuda_graph_warmup_steps = 3
+    cfg.model.use_te_rng_tracker = True
+    cfg.rng.te_rng_tracker = True
+
+    cfg.model.moe_token_dispatcher_type = "flex"
+    cfg.model.moe_flex_dispatcher_backend = "hybridep"
+    cfg.model.moe_flex_dispatcher_num_sms = 32
+    cfg.model.moe_hybridep_num_sms = None
+    cfg.model.moe_router_dtype = "fp32"
+    cfg.model.moe_shared_expert_overlap = False
+    cfg.model.moe_router_force_load_balancing = False
+    cfg.model.moe_router_padding_for_fp8 = False
+
+    cfg.optimizer.use_precision_aware_optimizer = False
+    cfg.optimizer.main_grads_dtype = torch.float32
+    cfg.optimizer.main_params_dtype = torch.float32
+    cfg.optimizer.exp_avg_dtype = torch.float32
+    cfg.optimizer.exp_avg_sq_dtype = torch.float32
+    cfg.optimizer.overlap_param_gather_with_optimizer_step = False
+
+    cfg.ddp.overlap_grad_reduce = False
+    cfg.ddp.overlap_param_gather = False
+    cfg.ddp.check_for_nan_in_grad = True
+    cfg.ddp.use_distributed_optimizer = True
+    cfg.ddp.use_megatron_fsdp = False
+
+    cfg.comm_overlap = CommOverlapConfig(
+        tp_comm_overlap=True,
+        overlap_grad_reduce=False,
+        overlap_param_gather=False,
+    )
+    return cfg
diff --git a/src/megatron/bridge/recipes/qwen/qwen35.py b/src/megatron/bridge/recipes/qwen/qwen35.py
new file mode 100644
index 0000000000..51025d0074
--- /dev/null
+++ b/src/megatron/bridge/recipes/qwen/qwen35.py
@@ -0,0 +1,27 @@
+# Copyright (c) 2026, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ruff: noqa: F401
+"""Compatibility alias for the canonical Qwen3.5 text recipe."""
+
+from __future__ import annotations
+
+from megatron.bridge.recipes.qwen.gb200.qwen35 import (
+    qwen35_text_9b_pretrain_8gpu_gb200_bf16_config as qwen35_text_9b_pretrain_config,
+)
+from megatron.bridge.recipes.qwen.gb200.qwen35 import (
+    qwen35_text_35b_a3b_pretrain_8gpu_gb200_bf16_config as qwen35_text_35b_a3b_pretrain_config,
+)
+
+
+__all__ = ["qwen35_text_9b_pretrain_config", "qwen35_text_35b_a3b_pretrain_config"]
diff --git a/tests/unit_tests/recipes/recipe_test_utils.py b/tests/unit_tests/recipes/recipe_test_utils.py
index b8c483c241..334062bba9 100644
--- a/tests/unit_tests/recipes/recipe_test_utils.py
+++ b/tests/unit_tests/recipes/recipe_test_utils.py
@@ -152,6 +152,11 @@ def finalize(self) -> None:
 class _OfflineAutoBridge:
     """Build a local provider without reading a Hugging Face configuration."""
 
+    @classmethod
+    def from_hf_config(cls, *args: object, **kwargs: object) -> "_OfflineAutoBridge":
+        del args, kwargs
+        return cls()
+
     @classmethod
     def from_hf_pretrained(cls, *args: object, **kwargs: object) -> "_OfflineAutoBridge":
         del args, kwargs
@@ -215,7 +220,17 @@ def skip_flex_dispatcher_hardware_probe(*args: object, **kwargs: object) -> None
     deepseek_v4_recipe_module = importlib.import_module("megatron.bridge.recipes.deepseek.h100.deepseek_v4")
     monkeypatch.setattr(deepseek_v4_recipe_module, "deepseek_v4_supports_blackwell_fused_kernels", lambda: False)
 
-    from transformers import AutoTokenizer
+    from transformers import AutoConfig, AutoTokenizer
+
+    def load_offline_auto_config(*args: object, **kwargs: object) -> SimpleNamespace:
+        del args, kwargs
+        return SimpleNamespace(text_config=SimpleNamespace(architectures=None))
+
+    monkeypatch.setattr(
+        AutoConfig,
+        "from_pretrained",
+        staticmethod(load_offline_auto_config),
+    )
 
     def load_offline_tokenizer(*args: object, **kwargs: object) -> _OfflineTokenizer:
         del args, kwargs
diff --git a/tests/unit_tests/recipes/test_nemotronh_recipes.py b/tests/unit_tests/recipes/test_nemotronh_recipes.py
index 98177dbdc9..c60a18cfd2 100644
--- a/tests/unit_tests/recipes/test_nemotronh_recipes.py
+++ b/tests/unit_tests/recipes/test_nemotronh_recipes.py
@@ -59,8 +59,9 @@ def to_megatron_provider(self, *args, **kwargs):
 
 @pytest.fixture(autouse=True)
 def _patch_hf_backed_recipe_providers(monkeypatch: pytest.MonkeyPatch) -> None:
-    """Keep Super and Ultra recipe construction deterministic and offline."""
+    """Keep AutoBridge-backed recipe construction deterministic and offline."""
     for module_name in (
+        "megatron.bridge.recipes.nemotronh.gb200.nemotron_3_nano",
         "megatron.bridge.recipes.nemotronh.nemotron_3_super",
         "megatron.bridge.recipes.nemotronh.nemotron_3_ultra",
     ):
@@ -138,6 +139,13 @@ def test_nemotronh_recipe_rejects_unknown_cli_override():
     assert not hasattr(cfg.model, "not_a_real_field")
 
 
+def test_nemotron_3_nano_gb200_defers_vocab_size_to_training_tokenizer():
+    """The GB200 pretraining model vocabulary must follow its runtime tokenizer."""
+    cfg = _nemotronh_module.nemotron_3_nano_pretrain_8gpu_gb200_bf16_config()
+
+    assert cfg.model.vocab_size is None
+
+
 def test_nemotron_nano_9b_v2_lora_defaults():
     """Test that Nemotron Nano 9B v2 LoRA has correct default parallelism."""
     from megatron.bridge.recipes.nemotronh import nemotron_nano_9b_v2_peft_config
diff --git a/tests/unit_tests/recipes/test_qwen_recipes.py b/tests/unit_tests/recipes/test_qwen_recipes.py
index 6bcb9a39cd..5f965b65c7 100644
--- a/tests/unit_tests/recipes/test_qwen_recipes.py
+++ b/tests/unit_tests/recipes/test_qwen_recipes.py
@@ -77,6 +77,26 @@ def from_hf_pretrained(hf_path: str, **kwargs):
             assert kwargs == {"revision": expected_revisions[hf_path]}
         return _FakeBridge()
 
+    @staticmethod
+    def from_hf_config(hf_config):
+        # Ignore hf_config; return a bridge that yields a fake provider
+        return _FakeBridge()
+
+
+class _FakeTextConfig:
+    architectures = None
+
+
+class _FakeRootConfig:
+    text_config = _FakeTextConfig()
+
+
+class _FakeAutoConfig:
+    @staticmethod
+    def from_pretrained(hf_path: str):
+        # Ignore hf_path; return a unified config with a nested text config.
+        return _FakeRootConfig()
+
 
 def _assert_basic_config(cfg):
     from megatron.bridge.training.config import ConfigContainer
@@ -114,6 +134,8 @@ def test_each_qwen_recipe_builds_config(recipe_func: Callable, monkeypatch: pyte
     mod = importlib.import_module(module_name)
     if hasattr(mod, "AutoBridge"):
         patch_recipe_module_global(monkeypatch, mod, "AutoBridge", _FakeBridge)
+    if hasattr(mod, "AutoConfig"):
+        patch_recipe_module_global(monkeypatch, mod, "AutoConfig", _FakeAutoConfig)
 
     overrides = _safe_overrides_for(recipe_func.__name__)
 
@@ -150,7 +172,12 @@ def test_each_qwen_recipe_builds_config(recipe_func: Callable, monkeypatch: pyte
     assert getattr(cfg.model, "tensor_model_parallel_size", 1) >= 1
     assert getattr(cfg.model, "pipeline_model_parallel_size", 1) >= 1
 
-    if "qwen3" in recipe_name and "pretrain" in recipe_name and "next" not in recipe_name:
+    if (
+        "qwen3" in recipe_name
+        and "pretrain" in recipe_name
+        and "next" not in recipe_name
+        and "qwen35" not in recipe_name
+    ):
         assert cfg.model.cross_entropy_fusion_impl == "te"
 
     # SFT and PEFT-specific assertions