nemotron-3-ultra-nvfp4

Nemotron 3 Ultra NVFP4 is NVIDIA's largest open-weight reasoning model in the Nemotron family, distinguished by a hybrid Mamba-Transformer LatentMoE design that activates 55 billion of its 550 billion parameters per token. Multi-Token Prediction layers enable native speculative decoding, and the model is pre-trained directly in NVIDIA's 4-bit NVFP4 floating-point format with a quantization-aware recipe that keeps most layers compact while reserving select layers at higher precision for stability. The NVFP4 build is complemented by a BF16 sibling for organizations whose hardware configurations favor standard precision, giving teams a choice between memory efficiency and conventional deployment paths.

The model targets the most demanding agentic workloads: long-context analysis, tool use, and high-accuracy reasoning across code, mathematics, and science. It supports context lengths up to one million tokens in the broader NVIDIA release, making it well suited to multi-step agents that need to hold entire codebases, research papers, or operational logs in working memory. NVFP4 quantization reduces accelerator memory pressure, and the build is supported on NVIDIA Hopper and Blackwell GPUs with mainstream inference frameworks including vLLM, SGLang, and TensorRT-LLM. Released under the permissive OpenMDW License v1.1, the model is a practical fit for teams building complex reasoning pipelines that need frontier-scale capacity without the cost profile of a fully dense checkpoint.