What is Nemotron 3 Ultra?
Nemotron 3 Ultra is NVIDIA's largest open language model and the biggest open-weight release from a US company, published on Hugging Face on 4 June 2026. It is a mixture-of-experts model with 550 billion parameters, 55 billion active per token, that mixes Mamba-2 layers (a fast alternative to attention for long sequences) with a few attention layers and adds multi-token prediction for faster output. It was pretrained in 4-bit NVFP4 precision, accepts up to 1M tokens of context, and reasoning can be switched on or off. It reads and writes text in ten languages.
The weights are under OpenMDW-1.1, which allows commercial use, and NVIDIA also publishes the pre-training and post-training datasets and recipes. The BF16 checkpoint needs eight B200, GB200 or H200 GPUs or sixteen H100s; an NVFP4 version targets Blackwell. On Hugging Face the BF16 repository has 383 likes and 1,122,228 downloads in the last 30 days.
On OpenRouter, DeepInfra serves it from $0.40 per million input tokens and $1.76 per million output tokens, with a $0.50/$2.20 tier, Baseten at $0.60/$2.40 and Venice at $0.625/$3.125.
How much does Nemotron 3 Ultra cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | OpenMDW-1.1, training data published |
| DeepInfra | $0.40 in / $1.76 out per 1M tokens | Also a $0.50/$2.20 endpoint |
| Baseten | $0.60 in / $2.40 out per 1M tokens | |
| Venice | $0.625 in / $3.125 out per 1M tokens |
Free weights under OpenMDW-1.1. On OpenRouter: DeepInfra $0.40/$1.76 or $0.50/$2.20, Baseten $0.60/$2.40, Venice $0.625/$3.125 per 1M input/output tokens.
Is Nemotron 3 Ultra worth it?
Nemotron 3 Ultra is a capable reasoning and agent model whose real selling point is openness. At launch Artificial Analysis scored it 47.7 on its Intelligence Index v4.0, ahead of Gemma 4 31B (39.2) and gpt-oss-120b (33.3), which made it the top US open model, but behind Kimi K2.6 (53.9). NVIDIA also publishes its training data and recipes, so you can see what went into it.
Its best traits are speed and restraint. Artificial Analysis measured over 400 output tokens a second on a pre-release deployment, faster than gpt-oss-120b at more than four times the size, and 71% on its non-hallucination test, meaning it tends to decline rather than invent an answer. NVIDIA's own table gives 78.7 on OmniScience non-hallucination against 5.7 for DeepSeek-V4-Pro, and 94.7 on 1M-token RULER.
The catch: it is not a top coder or a top knowledge model. NVIDIA reports 56.4 on Terminal-Bench 2.1 against 67.2 for Kimi K2.6, and 26.7 on Humanity's Last Exam against 37.7 for DeepSeek-V4-Pro; Poolside's 118B Laguna S 2.1 scores 70.2 on the same Terminal-Bench. Hosted context is capped below the 1M the model supports, at 262,144 tokens on DeepInfra with 16,384 output tokens.
Our take: Nemotron 3 Ultra is what you reach for when the auditor asks where the model came from. For pure capability per dollar, GLM-5.3 and Kimi K3 still lead, but neither comes with its training data.
Where it earns its place
A US healthcare or defence supplier with model-provenance rules
Deploy a frontier-class open model whose training data can be audited
Self-host the NVFP4 weights on eight Blackwell GPUs and hand reviewers NVIDIA's published pre-training and post-training datasets. OpenMDW-1.1 permits commercial use.
A research desk answering analyst questions over long reports
Get answers that say when the documents do not contain the fact
Call it on DeepInfra at $0.50/$2.20 with retrieval. Its low hallucination rate suits this work: 1,000 questions of 100K input and 2K output tokens cost about $54.
Before you commit
- Setup
- Call nvidia/nemotron-3-ultra-550b-a55b on OpenRouter, DeepInfra, Baseten or Venice, or download BF16 or NVFP4 weights; reasoning is toggled with enable_thinking in the chat template.
- Integrations
- OpenAI-compatible APIs; vLLM, SGLang and TensorRT-LLM; NVIDIA's NeMo Evaluator recipes reproduce its benchmark numbers.
- Data and privacy
- Open weights plus published training datasets; pre-training data runs to September 2025 and post-training data to May 2026. Self-hosting keeps prompts in-house; OpenRouter says EU in-region routing is not available for it.
- Cost at scale
- From $0.40/$1.76 per million tokens on DeepInfra. Self-hosting BF16 needs eight B200, GB200 or H200 GPUs, or sixteen H100s.
- Limits
- Up to 1M tokens of context in the model; hosted endpoints offer 202,800 to 262,144 tokens, and DeepInfra caps output at 16,384 tokens.
What’s good
- Training data and recipes published alongside the weights
- 47.7 on Artificial Analysis's index at launch, top among US open models
- Over 400 output tokens a second on a pre-release deployment
- Low hallucination: 71% on Artificial Analysis's non-hallucination test
What’s not
- Behind Kimi K2.6 and DeepSeek-V4-Pro on coding and hard knowledge, per NVIDIA's table
- Hosted endpoints cap context at 262,144 tokens or less
- BF16 needs eight Blackwell or H200 GPUs, or sixteen H100s
- Short 16,384-token output limit on DeepInfra
Best for
Companies that need a large, fast open reasoning model with documented training data, hosted or on their own GPUs.
Nemotron 3 Ultra alternatives
Nemotron 3 Ultra: common questions
- Is Nemotron 3 Ultra worth it?
- As open as a large model gets, quick and reluctant to make things up, though China's open leaders still out-reason it. We rate it 4.1 out of 5.
- How much does Nemotron 3 Ultra cost?
- Free weights under OpenMDW-1.1. On OpenRouter: DeepInfra $0.40/$1.76 or $0.50/$2.20, Baseten $0.60/$2.40, Venice $0.625/$3.125 per 1M input/output tokens.
- What are the downsides of Nemotron 3 Ultra?
- Behind Kimi K2.6 and DeepSeek-V4-Pro on coding and hard knowledge, per NVIDIA's table; hosted endpoints cap context at 262,144 tokens or less; bF16 needs eight Blackwell or H200 GPUs, or sixteen H100s; short 16,384-token output limit on DeepInfra.
- Who is Nemotron 3 Ultra best for?
- Companies that need a large, fast open reasoning model with documented training data, hosted or on their own GPUs.
- How much does Nemotron 3 Ultra cost for a team or in production?
- From $0.40/$1.76 per million tokens on DeepInfra. Self-hosting BF16 needs eight B200, GB200 or H200 GPUs, or sixteen H100s.
- Is Nemotron 3 Ultra safe for company data?
- Open weights plus published training datasets; pre-training data runs to September 2025 and post-training data to May 2026. Self-hosting keeps prompts in-house; OpenRouter says EU in-region routing is not available for it.
- What are the best Nemotron 3 Ultra alternatives?
- The closest alternatives we have reviewed are Kimi K3, GLM-5.3, DeepSeek-V4-Pro and Nemotron 3.5 Super VL.
Answers checked Nov 7, 2026.
Make Nemotron 3 Ultra? Show this rating on your site
