Qwen#

This is the landing page for Qwen model guidance in NeMo RL. It links to version-specific subpages and points to the recipes and known issues for each supported variant.

For the full list of supported Qwen models, see Model Support.

Version Guides#

  • Qwen3.5 — LLM and VLM recipes for Qwen3.5-9B-Base, Qwen3.5-35B-A3B-Base, and Qwen3.5-397B-A17B on the Megatron and AutoModel backends. This guide covers backend and parallelism support, example recipes, and the flash-linear-attention performance notes.

Subpages for other Qwen versions are added as distinct, recipe-backed guidance accumulates. Until then, the GRPO, evaluation, and recipe YAML files remain the source of truth for those models.

Note

Qwen3 and Qwen3.5 thinking models need a large generation budget. See Qwen3.5 → Example Recipes for the max_new_tokens guidance.

On this page