DeepSeek#

This section contains guides for running DeepSeek models with NeMo RL. For the complete support matrix, see Model Support.

Version Guides#

  • DeepSeek V3 — Prepare a BF16 Hugging Face checkpoint for DeepSeek V3.

  • DeepSeek V4 Flash — Run GRPO with DeepSeek V4 Flash or Flash Base using the AutoModel training backend and vLLM generation.

On this page