NVIDIANemotron 3
Nemotron 3 Ultra 550B-A55B
Ultra is NVIDIA's 550B-total, 55B-active hybrid LatentMoE reasoning model. It is a released checkpoint, not just an announced family size.
What the release provides
The card supports context up to one million tokens. The architecture mixes Mamba-2, experts and attention, with multi-token prediction. A chat-template flag controls reasoning behavior.Nemotron 3 Ultra 550B-A55B official model card
How to use it well
Practical guidance. Use a model of this size when the quality gain justifies the serving system it requires. Compare it with smaller models on complete workflows, including retrieval and tool failures. Budget for all stored experts and the desired context and concurrency.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 |
| Release license | OpenMDW 1.1Nemotron 3 Ultra 550B-A55B official model card |