Nemotron 3 NanoOpen Source

Nemotron 3 Nano, released December 15, 2025, opened NVIDIA's first fully in-house Nemotron generation — no Llama base underneath — and made the architectural bet explicit: a hybrid of Mamba-2 state-space layers, mixture-of-experts feed-forwards and a small number of attention layers, with 30 billion total parameters and roughly 3.5 billion active per token. The design traded a little per-token capacity for throughput on long agentic transcripts, which is what NVIDIA said models would spend 2026 doing.

Its launch card reported 73.0% on GPQA Diamond, 68.3% on LiveCodeBench and 89.1% on AIME 2025 without tools — unusually strong mathematics for the size — and the release was open in the fullest sense the industry used: weights, the bulk of the datasets and the training recipes all published. Super and Ultra tiers were promised for 2026 and arrived in March and June of that year.