Chat with Nemotron 3 Ultra online

Nemotron 3 Ultra is NVIDIA's open frontier model, released on 4 June 2026 after being announced at Computex. It is a 550-billion-parameter mixture-of-experts model with about 55 billion parameters active per token, and unusually it is not a plain transformer: NVIDIA built it on a hybrid Transformer-Mamba architecture, which is part of how it sustains high throughput on long inputs.

NVIDIA is explicit about what it is for. This is a reasoning and orchestration model, aimed at long-running agentic work — an agent that coordinates other agents, a coding agent that keeps going, deep research that spans many steps, and enterprise tasks that do not fit in a single prompt. The design includes multi-token prediction layers with shared weights, which improves the training signal and enables native speculative decoding, so it produces tokens quickly for a model of its size.

It is genuinely open. The weights ship under the OpenMDW 1.1 licence for commercial and non-commercial use, and NVIDIA published training recipes and a multi-trillion-token pre-training dataset alongside them, which is a level of disclosure most labs at this scale do not offer. NVIDIA documents a pre-training data cutoff of September 2025 with post-training data running to May 2026.

On NinjaChat it serves about 256K tokens of context with up to 128K tokens of output — the shape our providers run — so you can hand it a long research task or a multi-stage plan and let it work.