NVIDIA Nemotron 3 Ultra

About

NVIDIA presents a 550 billion parameter Mixture-of-Experts model, which utilizes a hybrid Mamba-Attention design. It achieves over 300 tokens per second and supports a context window of one million tokens. Ranked as the leading open-weights model in the United States on the Artificial Analysis Intelligence Index, it is engineered for complex multi-step agent workflows that require advanced, cost-effective reasoning. The model is currently accessible on Hugging Face, OpenRouter, ModelScope, and as a NVIDIA NIM microservice at build.nvidia.com.