Available on request
Nemotron API:
open weights, open data, open recipes.
NVIDIA publishes Nemotron with its weights, training data and recipes. Nemotron 3 Ultra is the largest model in the family, built for reasoning and agent work over long context. Tell us your volume, and we’ll talk through adding it to the Xinference Model API.
Under the hood
A hybrid of Mamba-2, attention and experts
Most large open models are pure transformers. Nemotron 3 Ultra interleaves Mamba-2 layers with Mixture-of-Experts layers and a few attention layers, a design NVIDIA calls LatentMoE. Mamba-style layers handle long sequences more cheaply than attention alone, which is part of how the model reaches a 1M-token context.
| Model | Nemotron-3-Ultra-550B-A55B |
| Parameters | 550B total, 55B active |
| Context | Up to 1M tokens Model maker’s figure |
| Reasoning | Switch on or off per request (enable_thinking) |
| Released | 4 June 2026 |
| Licence | OpenMDW 1.1 |
Source: NVIDIA’s Nemotron 3 Ultra model card, checked 23 October 2026. Nemotron, NVIDIA and the NVIDIA logo are trademarks of NVIDIA Corporation, shown only to identify the model. Their use does not imply endorsement.
Ten languages
Built for multilingual reasoning
Supported languages
English, French, Spanish, Italian, German, Japanese, Hindi, Korean, Brazilian Portuguese and Chinese.
Where NVIDIA points it
Frontier reasoning, complex agentic workflows, long-context analysis, tool use and high-stakes retrieval-augmented generation.
Nemotron API questions
What is the OpenMDW licence?
OpenMDW 1.1 is a permissive licence written for models. It allows use, modification and distribution without restriction. If you redistribute the model, you keep a copy of the licence and the original notices.
Can I turn Nemotron’s reasoning off?
Yes. NVIDIA makes reasoning configurable through the chat template, so a request can ask for step-by-step reasoning or a direct answer.
Is Nemotron on the Xinference Model API?
On request. The Model API’s standard models are Kimi K3 and DeepSeek V4.1 Flash, with public per-token pricing. For Nemotron, tell us your expected volume.
Do you train on my data?
No. We don’t use your prompts or outputs to train or fine-tune models unless you agree in writing.
Last updated