Nemotron 3 Nano
NVIDIA's 30B-A3B hybrid Mamba-MoE open model — 1M context, open weights plus training recipes.
About
Nemotron 3 Nano is the first release of NVIDIA's Nemotron 3 family (announced December 15, 2025; Super and Ultra followed in 2026). It is a 30B-total / 3B-active-parameter model using a hybrid Mamba-2 + Transformer mixture-of-experts architecture trained with NVFP4, multi-environment reinforcement learning, and inference-time reasoning budget control. Context runs to 1M tokens (262K on the Hugging Face build).
NVIDIA released the weights under the NVIDIA Open Model License along with its synthetic pretraining corpus (~10T tokens), training recipes, and RL environments — a genuine reproducibility play, not just a weights drop. Independent signal: it leads the Artificial Analysis Intelligence Index v3 (score 52) among similarly-sized models and preserved the AI Openness Index score its Nano V2 predecessor held. It is available on Hugging Face, NVIDIA NIM, and build.nvidia.com, and is designed for high-throughput agentic and multi-agent workloads on H100/B200-class hardware.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Independent (Artificial Analysis Intelligence Index v3): 52, leading among similarly-sized models.
Provider-reported: best-in-class inference throughput for its accuracy tier.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Verified against NVIDIA's official developer blog, the published technical white paper, and independent press (The Register); independent Artificial Analysis index corroborates the accuracy claims. Open weights, data, and recipes are downloadable and checkable. Docked only for: (1) throughput superlatives are provider-measured; (2) the NVIDIA Open Model License is permissive but not OSI open source — check commercial terms.