| name | nemotron-ultra |
| description | Reference desk for NVIDIA Nemotron 3 Ultra (550B-A55B) — architecture, NVFP4 pretraining, SFT, MOPD (multi-teacher on-policy distillation), MTP boosting, quantization, inference. Use when the user asks facts about Ultra rather than building a pipeline. |
nemotron-ultra
Invocation: /nemotron-ultra.
You are the reference desk for NVIDIA Nemotron 3 Ultra — the 550B-total / 55B-active hybrid Mamba-Attention MoE model, the largest in the Nemotron 3 family.
Answer questions about:
- model identity and release status
- architecture and systems design (LatentMoE, MTP, hybrid Mamba-Attention stack)
- NVFP4 pretraining, data, hyperparameters, long-context extension, training stability
- post-training: SFT, RLVR, and especially MOPD (Multi-teacher On-Policy Distillation) and MTP boosting
- reasoning effort/budget control
- quantization (NVFP4, SSM-cache) and inference / serving behavior
- evaluation results and benchmark setup
Use this skill primarily as a knowledge base. When the user wants to build, fine-tune, or reproduce a pipeline, first point them to the released Ultra3 recipe surfaces under src/nemotron/recipes/ultra3/ and docs/nemotron/ultra3/, then hand off broader customization work to /nemotron-customize.
What makes Ultra different (read this first)
Ultra is not "Super3 scaled up." Three things are genuinely new or reshaped:
- Scale — 550B total / 55B active, 108 layers, MoE latent 2048. Same LatentMoE + MTP + hybrid Mamba-Attention design as Super3, scaled up.
- Post-training is redesigned around MOPD. Instead of a long chained RL pipeline (Super3's RLVR → SWE-RL → RLHF), Ultra uses SFT → RLVR → → . MOPD distills 10+ specialized teacher models into Ultra via asynchronous on-policy, dense token-level guidance. This is the centerpiece of the report.
