Nemotron#

This is the landing page for Nemotron model guidance in NeMo RL. It links to model-specific subpages that cover the released post-training recipes, launch instructions, and known issues for each variant.

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

Model Guides#

  • Nemotron 3 Nano — GRPO math post-training for the 30B-A3B hybrid Mamba MoE model, including data preparation and Slurm launch scripts.

  • Nemotron 3 Nano Omni — GRPO for the Nano Omni vision-language model on the AutoModel and Megatron backends (CLEVR-CoGenT and MMPR-Tiny recipes).

  • Nemotron 3 Super — the multi-stage Nemotron 3 Super post-training recipe (RLVR, SWE, and RLHF stages).

  • Nemotron 3 Super Omni Image MOPD — image on-policy distillation for the Super Omni vision-language model with a non-colocated teacher (10-node production recipe plus a 4-node smoke).

  • Nemotron 3 Ultra — RLVR, teacher training, and MOPD stages on GB200 NVL72 hardware.

  • Nemotron 3.5 Lightning — RLVR with NeMo Gym on GB200, plus a compact 4-node DAPO math recipe on the DTensor (AutoModel) backend.

On this page