Nemotron 3 Ultra
NVIDIA's largest open model: 550B hybrid Mamba-Transformer MoE with 55B active, 1M context, trained in NVFP4, under the OpenMDW license.
In short. Nemotron 3 Ultra is the biggest open model NVIDIA has released. It is built for hard reasoning and long agent tasks, reads up to 1 million tokens, and speaks ten languages. It needs a server with 8 to 16 top-end GPUs, but the license is permissive and allows commercial use.
Summary
NVIDIA released Nemotron 3 Ultra on 4 June 2026 as the largest model of the Nemotron 3 family. It has 550 billion parameters, of which 55 billion are active per token. The architecture interleaves Mamba-2 state-space layers, mixture-of-experts layers and a few attention layers. It uses LatentMoE, which routes tokens through a smaller latent space, and multi-token-prediction layers for faster generation. It was pretrained on about 20 trillion tokens with an NVFP4 recipe, in which most linear layers use 4-bit floating point for weights, activations and gradients. The context window is up to 1 million tokens, and the model supports English, French, Spanish, Italian, German, Japanese, Hindi, Korean, Portuguese and Chinese.
NVIDIA's table compares Ultra with MiniMax M2.7, GLM-5.1, Kimi K2.6, Qwen3.5 and the DeepSeek V4 models. It is mixed: Ultra scores 94.7 on RULER at 1M tokens and leads that row, but on SWE-Bench Verified it gets 70.7 against 75.7 for Kimi K2.6 and 76.2 for GLM-5.1, and on Terminal Bench 2.1 it gets 56.4 against 67.2 for Kimi. These are NVIDIA's own results. Use the card's numbers as a guide and test on your own tasks.
The BF16 checkpoint is about 1.1 TB and the NVFP4 one about 352 GB. The card lists 8 Blackwell GPUs (B200, GB200, B300 or GB300), 8 H200 or 16 H100 as the minimum. Unlike the Nemotron 3 Nano and Super models, which use the NVIDIA Nemotron Open Model License, Ultra uses OpenMDW-1.1, a permissive license that covers model, data and weights and allows commercial and non-commercial use.
Good for
- Frontier-class agent, coding and long-document work on your own servers
- Multilingual reasoning across ten languages
- Organizations that want a permissive license on a very large model
- Research on NVFP4 pretraining and LatentMoE at scale
Watch out for
- It needs 8 to 16 data-center GPUs even in NVFP4, which is about 352 GB
- On coding and terminal benchmarks, NVIDIA's own table shows it behind GLM-5.1 and Kimi K2.6
- It is text-only; use Nemotron 3.5 Super VL for images and video
- The benchmark numbers come from NVIDIA
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the Ultra 550B-A55B (BF16) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| Ultra 550B-A55B (BF16) BF16 | 550B 55B active | 1.26 TB | 633 GB | 379 GB | Big-memory desktop |
| Ultra 550B-A55B (NVFP4) NVFP4 | 550B 55B active | 1.26 TB | 633 GB | 387 GB | Big-memory desktop |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
A short permissive license written for machine-learning models by the Linux Foundation and PyTorch Foundation. It covers the model, its data, documentation and software together, and puts no restrictions on what you do with the model's outputs.
See also
Frequently asked questions
- What is Nemotron 3 Ultra?
- Nemotron 3 Ultra is the biggest open model NVIDIA has released. It is built for hard reasoning and long agent tasks, reads up to 1 million tokens, and speaks ten languages. It needs a server with 8 to 16 top-end GPUs, but the license is permissive and allows commercial use.
- Can I use Nemotron 3 Ultra commercially?
- Yes. Nemotron 3 Ultra is released under OpenMDW-1.1, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Nemotron 3 Ultra open source?
- Nemotron 3 Ultra is open-weight, not open source in the strict sense. You can download the model files, but the OpenMDW-1.1 adds conditions that the Open Source Initiative definition does not allow.
- Can I run Nemotron 3 Ultra on my own computer?
- Only on a machine with a lot of memory. The smallest variant, Ultra 550B-A55B (BF16), needs about 379 GB at 4-bit precision. Desktops with 128 to 512 GB of unified memory can hold it, at modest speed.
- When was Nemotron 3 Ultra released?
- Nemotron 3 Ultra was released by NVIDIA on June 4, 2026.
- How long a context window does Nemotron 3 Ultra have?
- Nemotron 3 Ultra supports a context window of up to 1M tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Nemotron 3 Ultra 550B-A55B BF16 model card (Hugging Face) huggingface.co
- NVIDIA Nemotron 3 family of models (NVIDIA Research) research.nvidia.com
- NVIDIA Nemotron 3 Ultra Technical Report research.nvidia.com
- OpenMDW License Agreement, version 1.1 raw.githubusercontent.com
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Nemotron 3 NVIDIA | 31.6B – 120B 12B active | 1M | NVIDIA Nemotron (Permissive) | 21 GB Gaming GPU | |
| Nemotron 3.5 NVIDIA | 8B – 120B 12B active | 1M | OpenMDW-1.1 (Permissive) | 5.52 GB Laptop or phone | |
| DeepSeek V4 DeepSeek | 284B – 1.6T 49B active | 1M | MIT (Permissive) | 176 GB Big-memory desktop | |
| GLM-5.2 Z.ai (Zhipu AI) | 753B 40B active | 1M | MIT (Permissive) | 520 GB GPU server |