Nemotron 3 Nano (30B A3B)

Nemotron 3 Nano is a 31.6B hybrid MoE model optimized for fast, long‑context agentic reasoning. It mixes Mamba‑2 and Transformer layers with a sparse MoE router (~3.6B active params per token) to deliver up to 4× higher throughput than Nemotron 2 and strong accuracy across math, coding, and tools. It supports a 1M‑token context window, offers Reasoning ON/OFF and a thinking‑budget to control costs, and ships with open weights, data, and RL tooling (NeMo Gym/RL). Released Dec 15, 2025 under the NVIDIA Open Model License, it’s built as the efficient backbone for multi‑agent systems at scale.