Nvidia releases large open AI model for autonomous agents

Nvidia unveiled the Nemotron 3 Ultra model at the Computex trade show on June 1, 2026. It was released on June 4. The model is aimed at so-called agents, meaning AI systems that work independently over long periods, for example on coding and research.

The model is open, which means that both the weights (the numbers that make up the model's knowledge) and the data are available. The license is OpenMDW-1.1 from the Linux Foundation. It can be used through Hugging Face, build.nvidia.com, Perplexity, OpenRouter and more than 30 cloud and inference providers, among others.

The technical details

Nemotron 3 Ultra is a mixture-of-experts model. That means only part of the model is active at a time: about 55 billion of roughly 550 billion total parameters. It can handle up to 1 million tokens at once, equivalent to very long documents or workflows. The architecture is a blend of Mamba and Transformer, and the model uses the NVFP4 format, which according to Nvidia is compatible with Hopper, Blackwell and Ampere GPUs.

Nvidia says the model delivers up to 5 times higher throughput than other open models, and that the cost of completing tasks in the SWE-bench and Terminal-Bench 2.0 tests can fall by up to 30%. These are Nvidia's own claims. The company's own tests give 54% on Terminal-Bench 2.0 and 95% on Ruler at 1 million tokens, among other results.

Independent assessment

The company Artificial Analysis has measured the model at 47.7 on its Intelligence Index. That makes it the most intelligent American open model, but it trails Chinese open models such as Kimi K2.6, which scores 53.9. Artificial Analysis gained access through a partnership with Nvidia before the release.

On a pre-release version, it measured more than 400 output tokens per second. The model is more than four times larger than gpt-oss-120b, but according to the measurement it has comparable speed and higher intelligence. The model scores 71% on AA-Omniscience Non-Hallucination, which measures how well it avoids fabricated answers, and 3% on the physics test CritPt.

Other models and users

Also released were Nemotron 3.5 Content Safety, a safety model with 4 billion parameters, and Nemotron 3.5 ASR, a speech recognition model for more than 40 languages. Early users of the Nemotron family include Accenture, Deloitte, Oracle Cloud Infrastructure and Zoom. Perplexity, Palantir, ServiceNow and CrowdStrike are named specifically for Ultra.