Nvidia Nemotron 3 Ultra Prioritizes Speed for Agentic AI

Nvidia has launched Nemotron 3 Ultra, a large language model built on a hybrid transformer-mamba architecture, specifically designed for long-running agentic tasks. While its overall performance isn't top-tier, it is notably faster than comparable open-weight models, achieving around 183 tokens per second, three times faster than rivals like Moonshot Kimi K2.6. Nvidia also open-sourced its weights, training data, and reinforcement learning environments to encourage developer adoption.