Nemotron 3 Ultra

On June 4, 2026, NVIDIA dropped Nemotron 3 Ultra, and for once the "open" in "open model" is the whole point. It is a 550-billion-parameter model built for long-running agents, it tops Artificial Analysis as the most intelligent US open-weights model, and NVIDIA shipped not just the weights but the training data, the recipes, and the reinforcement-learning environments. It also did something quieter that matters more to builders: it moved off NVIDIA's restrictive model license onto a permissive one that allows commercial use without strings. Here is what it is, what is actually open, and why you still cannot run it at home.

What NVIDIA shipped, and which Nemotron this is

First, disambiguation, because NVIDIA's naming is a maze. Nemotron 3 Ultra is the June 2026 agentic flagship, the largest of the Nemotron 3 family (Nano, Super, Ultra). It is not Nemotron-Cascade 2, a separate and much smaller 30B model from March that was a post-training-recipe showcase. When people say "the new open NVIDIA model," they mean Ultra.

Ultra is a hybrid. Instead of a pure Transformer, it interleaves Mamba-2 state-space layers with attention and a Mixture-of-Experts design NVIDIA calls LatentMoE. (Mamba layers process long sequences in linear time rather than attention's quadratic cost, which is how you get a sane 1M-token context.) The shape: 550B total, 55B active per token, 512 experts with the top 22 routed, 108 layers, and native Multi-Token Prediction for speculative decoding. It was pretrained directly in 4-bit NVFP4 on about 20 trillion tokens. Text only, 12 languages.

The part that is actually open

Plenty of "open" models give you a weights file and nothing else. Nemotron 3 Ultra ships four checkpoints (base, post-trained, NVFP4, and a reward model), plus the pretraining data, the post-training recipes, and the RL environments, all on GitHub under NVIDIA-NeMo. You can reproduce the post-training, not just run the result.

And the license changed. Older Nemotron models used the NVIDIA Open Model License, which carried restrictions. Ultra ships under OpenMDW-1.1, a permissive, MIT-style license that grants commercial and non-commercial use, allows redistribution, and puts no restrictions on the model's outputs. For a company that has historically been careful with its licenses, that is a real shift toward genuine openness.

Benchmarks: the strongest US open model, with caveats

NVIDIA's own numbers, from the technical report, put Ultra firmly in the agentic mix. Independent confirmation comes from Artificial Analysis, which scored it at 48 on its Intelligence Index and called it the most intelligent US open-weights model as of June 2026. The honest footnote: Chinese open models still lead the overall open frontier (GLM-5.2 and Kimi score higher), so "best US open model" is the accurate claim, not "best open model."

NVIDIA-reported (technical report); Ultra run with TensorRT-LLM. The Intelligence Index ranking is independent (Artificial Analysis). Not reproduced by us.

Where it leads is long-context and agent throughput: RULER at 1M context (94.7) tops its peers, and NVIDIA claims roughly 5 to 6 times the throughput of GLM-5.1 and Kimi K2.6 on long agentic runs, the payoff of the Mamba hybrid and native NVFP4. Where it trails: raw coding-agent scores like Terminal-Bench, where Kimi sits higher.

Limitations and gotchas

• Not local. 550B in BF16 needs 8x H200 or 16x H100; even 4-bit NVFP4 wants 4x B200 or 8x H100. There is no consumer-GPU or Apple Silicon path.
• Best on Blackwell. NVFP4 is native on Blackwell; on older cards you fall back to a W4A16 path, which works but gives up part of the speed story.
• Text only. No vision or audio in Ultra; NVIDIA ships separate ASR and safety models.
• Benchmarks are NVIDIA-reported. The independent Artificial Analysis index corroborates the tier, not every number.

Who should use it

Use it if you are building long-horizon agents (multi-step planning, tool use, sub-agent delegation across hundreds of turns) and you want a fully open, commercially usable base you can fine-tune and self-host on a Blackwell or Hopper node. It integrates with CrewAI and LangChain agent stacks out of the box. If you want something you can run on your own GPU, this is not it; reach for a smaller Nemotron (Nano or Super) or a mid-size MoE. The reason to care about Ultra even if you cannot run it: the open data and recipes are a gift to anyone training their own agent model.

Run it in about 10 minutes

Realistically that means a hosted endpoint or a rented multi-GPU box. The fastest taste is the NVIDIA NIM playground or OpenRouter; the self-host path is vLLM.

# Hosted: try it free on OpenRouter or build.nvidia.com first.

# Self-host NVFP4 on 8x H100 (or 4x B200) with vLLM
docker run --gpus all vllm/vllm-openai:v0.22.0 \
  --model nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4 \
  --tensor-parallel-size 8

Then point a CrewAI or LangChain Deep Agents loop at the endpoint and hand it a task that needs planning and tool calls across many turns. That is the workload Ultra was post-trained for, and where the throughput advantage shows up. If you train agent models yourself, the real ten-minute move is reading the post-training recipes on GitHub, that is the part nobody else open-sourced.

Sources and further reading

• Nemotron 3 Ultra model card (Hugging Face)
• NVIDIA developer blog
• Nemotron 3 Ultra technical report (PDF)
• Open data, recipes, and RL environments (GitHub)
• Artificial Analysis: independent evaluation

Tested on: not independently tested. Nemotron 3 Ultra is a 550B hybrid MoE that needs a Blackwell or Hopper multi-GPU node even at NVFP4, beyond our bench. Benchmarks are NVIDIA-reported; the open-weights ranking is from Artificial Analysis, flagged as third-party. Sources linked above.
Date checked: 2026-06-26