Teams wanting a very large open agent model can prototype free on NVIDIA's endpoint, but self-hosting 550 billion parameters is no small job.
Kenning's take
What it does
Nemotron 3 Ultra is the biggest model in NVIDIA's Nemotron 3 family, with 550 billion total parameters and 55 billion active per token in a mixture-of-experts design. It is tuned for agent work, coding, planning and tool calls, and reads up to 1 million tokens of context. NVIDIA released the weights and training datasets, and offers a free trial endpoint and paid partner endpoints through its build.nvidia.com catalogue.
- 550B total, 55B active parameters
- Hybrid Mamba-Transformer mixture-of-experts
- Up to 1 million token context
- Reasoning budget control at inference time
- Open checkpoints and training datasets
- OpenAI-compatible API on build.nvidia.com
Who it suits. Developers and companies that want an open model they can host themselves for agent and coding work.
Key facts
- Category
- AI models
- Part of
- NVIDIA Nemotron 3
- Made by
- NVIDIA
- Based in
- United States
- Released
- 4 June 2026
- How to get it
- free API endpoint for prototyping on build.nvidia.com, partner endpoints, or download the weights to self-host
- Works on
- API, Self-hosted
- In the UK
- Yes
- Free plan
- free API endpoint for prototyping on build.nvidia.com and free weights
Reviews and write-ups
FAQ
When was Nemotron 3 Ultra released?
NVIDIA's research page for Nemotron 3 Ultra is dated 4 June 2026.
Is Nemotron 3 Ultra free?
The weights and training data are free to download, and build.nvidia.com offers a free endpoint for prototyping. Hosted partner endpoints charge their own prices.
How big is Nemotron 3 Ultra?
It has 550 billion total parameters, with 55 billion active per token, and a context window of up to 1 million tokens.
Facts from the maker's own pages and the sources listed. Checked 2026-10-11.
Similar tools
Nemotron 3 Nano Omni
An open NVIDIA model that reads video, audio, images and text together for AI agents
Nemotron 3 Super
NVIDIA's open 120B reasoning model with a 1 million token context, built for multi-agent systems
Nemotron 3.5 Lightning
A small, fast open NVIDIA model that handles the routine tool calls in long-running AI agents
ACE-Step 1.5
Open-source music generation model that writes full songs with lyrics and runs on ordinary consumer GPUs.