What is NVIDIA Nemotron 3 Ultra?
Nemotron 3 Ultra is the largest and final model in NVIDIA's Nemotron 3 family, above Nano and Super, and the most capable open model NVIDIA has released. It is a 550B mixture of experts that mixes Mamba-2 and attention layers, uses 55B parameters per token and reads up to 1M tokens of text. NVIDIA built it for the hard planning calls in agent systems, leaving routine steps to smaller models.
In NVIDIA's own table it scores below GLM-5.1 and Kimi K2.6 on coding but runs several times faster, and unlike most open models it ships with its training data and recipes. Its pre-training data stops at September 2025.
Who is NVIDIA Nemotron 3 Ultra best for?
NVIDIA Nemotron 3 Ultra is best for agent orchestration, very long documents and codebases, teams that must self-host an open model, and researchers who want the training data too.
Agent orchestration
NVIDIA trained it for the planning and judgment calls in long agent runs, and it scores 91% on PinchBench and 82% on IFBench in NVIDIA's tests.
Long documents and codebases
It keeps 95% accuracy on RULER at the full 1M tokens, a length GLM-5.1 and Kimi K2.6 cannot take.
Teams that must self-host
The weights are free for commercial use under OpenMDW-1.1, so prompts and data can stay on your own NVIDIA GPUs or in your cloud account.
Researchers and fine-tuners
NVIDIA released the base model, the reward model, much of the training data, and LoRA, SFT and reinforcement learning recipes.
What's included with NVIDIA Nemotron 3 Ultra?
Reasoning you can switch off or cap, multi-token prediction, one NVFP4 checkpoint for most NVIDIA GPUs, an OpenAI-compatible API, tool calling and ten languages.
- Reasoning you can control
- Turn thinking on or off, pick a medium-effort mode, or cap the reasoning trace with a token budget.
- Multi-token prediction
- Built-in layers predict several tokens per pass, which speeds up long answers through speculative decoding.
- One NVFP4 checkpoint
- The same 4-bit checkpoint runs on Ampere, Hopper and Blackwell GPUs; BF16 and base versions are also on Hugging Face.
- OpenAI-compatible API
- NVIDIA's endpoint takes the model ID nvidia/nemotron-3-ultra-550b-a55b, and the same model ships as a NIM container for your own servers.
- Tool calling
- Function calling works on NVIDIA's endpoint; structured output is listed as not supported there.
- Ten languages
- English, French, Spanish, Italian, German, Japanese, Hindi, Korean, Brazilian Portuguese and Chinese.
How much does NVIDIA Nemotron 3 Ultra cost?
Nemotron 3 Ultra is free: the weights download at no cost with commercial use allowed and NVIDIA runs a free trial API, while NVIDIA's NIM container needs an AI Enterprise licence at $4,500 per GPU per year in production.
- OpenMDW-1.1 licence, commercial use allowed
- NVFP4, BF16 and base checkpoints on Hugging Face
- You pay only for the GPUs to run it
- For prototyping, under NVIDIA's API trial terms
- Prompts and replies are recorded to improve NVIDIA products
- Needs a free NVIDIA API key
- Needed to run the NIM container in production
- 90-day free evaluation
- Education, Inception and Connect members pay $1,125
NVIDIA sells no per-token plan for Nemotron 3 Ultra. Partner hosts listed on NVIDIA's page set their own serverless prices: OpenRouter $0.60 / $2.40, GMI Cloud $0.80 / $2.60 and DigitalOcean $0.90 / $1.70 per million input / output tokens. NVIDIA AI Enterprise is also sold as a perpetual licence with five years of support at $22,500 per GPU, or at $1 per GPU per hour plus instance costs on cloud marketplaces. Full pricing on build.nvidia.com. We checked these prices on .
How do readers rate NVIDIA Nemotron 3 Ultra?
No one has rated NVIDIA Nemotron 3 Ultra yet. Be the first.
Sign in to rate. One rating per account.
NVIDIA Nemotron 3 Ultra pros and cons
It is fast for its size, fully open and strong at 1M-token inputs, but it needs a rack of data-centre GPUs, trails GLM-5.1 and Kimi K2.6 on coding, and reads text only.
Pros
- Fast for its size: NVIDIA measures 5.9x the throughput of GLM-5.1 and 4.8x that of Kimi K2.6, and early users single out its low latency.
- Fully open: the weights, much of the training data, the reinforcement learning environments and the recipes are public, which users say few large models offer.
- Built for very long inputs: 95% on RULER at 1M tokens in NVIDIA's table, and users of the same Nemotron design report it stays usable at deep context.
Cons
- Needs data-centre hardware: NVIDIA's minimum is 4 B200 or 8 H100 GPUs, and users found no practical way to run it on a 128GB home machine.
- Trails the strongest open models on coding: NVIDIA's own table has it at 54% on Terminal-Bench 2.0 against 64% for GLM-5.1 and 67% for Kimi K2.6, and several users rate it a step behind the best Chinese open models.
- Text only: it cannot read images, audio or video, and NVIDIA's endpoint does not support structured output.
Is NVIDIA Nemotron 3 Ultra open source?
Yes. NVIDIA Nemotron 3 Ultra's code is public under the OpenMDW 1.1 licence, so you can run it on your own server.
What are the best NVIDIA Nemotron 3 Ultra alternatives?
NVIDIA Nemotron 3 Ultra is an open source alternative to Nemotron 3 Super. Mistral Small 4 and Qwen3.6-27B do the same job and are open source too.
Nemotron 3 SuperFreeChoose it when you want NVIDIA's open 1M-context design on far less hardware; it has 12B active parameters out of 120B and needs only 2 H100 GPUs or one B200.
Mistral Small 4Open SourceMistral's 119B open MoE for chat, reasoning, coding and vision.
Qwen3.6-27BOpen SourceAlibaba's 27B open coding and vision model.
GLM-5.1Open SourceChoose it when coding accuracy matters more than speed or context; it scores 64% on Terminal-Bench 2.0 in NVIDIA's own table and costs $1.4 / $4.4 per million tokens on Z.ai, but its context stops at 200K.
Mistral 3Open SourceChoose it when you need image input and a plain Apache 2.0 licence; Mistral Large 3 is a 675B mixture of experts with 41B active that reads images and runs on one 8x H100 node.
Questions about NVIDIA Nemotron 3 Ultra
What is NVIDIA Nemotron 3 Ultra?
NVIDIA Nemotron 3 Ultra is NVIDIA's most capable open model, a free 550B mixture of experts with a 1M-token context, built to plan and reason inside long-running AI agents.
Who is NVIDIA Nemotron 3 Ultra for?
NVIDIA Nemotron 3 Ultra is best for agent orchestration, very long documents and codebases, teams that must self-host an open model, and researchers who want the training data too.
Is NVIDIA Nemotron 3 Ultra free?
Yes. NVIDIA Nemotron 3 Ultra has a free Open weights plan: openMDW-1.1 licence, commercial use allowed.
How much does NVIDIA Nemotron 3 Ultra cost?
NVIDIA Nemotron 3 Ultra's paid plans are NVIDIA AI Enterprise at $4,500 per GPU per year.
Is NVIDIA Nemotron 3 Ultra open source?
Yes. NVIDIA Nemotron 3 Ultra's code is public under the OpenMDW 1.1 licence, so you can run it on your own server.
What are the best NVIDIA Nemotron 3 Ultra alternatives?
The closest alternatives on Recaply are Nemotron 3 Super, Mistral Small 4, Qwen3.6-27B, GLM-5.1 and Mistral 3.
Where can I try NVIDIA Nemotron 3 Ultra?
On its official website, build.nvidia.com.
What hardware do I need to run Nemotron 3 Ultra?
NVIDIA's minimum for the NVFP4 weights is four B200, GB200, B300 or GB300 GPUs, or eight H100s. The default serving setups in vLLM, SGLang and TensorRT-LLM use a 256K context; the full 1M needs an extra setting.
Can I use Nemotron 3 Ultra commercially?
Yes. NVIDIA marks it ready for commercial and non-commercial use under OpenMDW-1.1, the Linux Foundation's licence for open models. The free hosted endpoint is for prototyping under NVIDIA's trial terms, and running NVIDIA's NIM container in production needs an AI Enterprise licence.
How does Nemotron 3 Ultra compare with GLM-5.1 and Kimi K2.6?
In NVIDIA's own table it leads both on IFBench and 1M-token recall, ties Kimi K2.6 at 91% on PinchBench, and trails both on Terminal-Bench 2.0 and GDPval-AA. NVIDIA says it runs 5.9x faster than GLM-5.1 and 4.8x faster than Kimi K2.6.
Can I turn off reasoning in Nemotron 3 Ultra?
Yes. Reasoning is on by default; set enable_thinking to false in the chat template for direct answers, add the medium-effort flag for shorter reasoning, or cap the reasoning trace with a token budget.
Does Nemotron 3 Ultra accept images?
No. It takes text in and gives text out. In the Nemotron 3 family, only the much smaller Nemotron 3 Nano Omni is multimodal, so pair Ultra with a separate vision model if your work needs both.
Our verdict
Nemotron 3 Ultra is the pick for teams that want a fast, fully open, frontier-sized model on their own NVIDIA GPUs.
It reads 1M tokens, ships with its training data and recipes, and runs several times faster than GLM-5.1 and Kimi K2.6 by NVIDIA's measure, though both score higher on coding.
- Skip it for
- Anyone without a multi-GPU server, anyone who needs image input, and anyone who wants the best open coding model regardless of speed.
- Check first
- Test it against GLM-5.1 on your own agent tasks, and price the GPUs or a partner host before you commit.
Nemotron 3
Nemotron 3 Nano Omni
Nemotron 3.5 Lightning
Ring-2.6-1T
Mistral Medium 3.5
DeepSeek-V3.2-Speciale