Overview
Nemotron 3 Nano was trained from scratch by NVIDIA and released in December 2025 as the first Nemotron 3 model. It combines Mamba-2 layers, mixture-of-experts layers (128 experts plus one shared, 6 active per token) and a few attention layers, and it produces a reasoning trace before its answer. NVIDIA also publishes much of the training data and recipes.
Highlights
- 30B total, about 3.5B active parameters
- Hybrid Mamba-2, MoE and attention architecture
- Unified reasoning and non-reasoning model
- Open weights, training data and recipes
- Ready for commercial use under NVIDIA’s open model license
Typical uses
Fast local agents, tool calling, coding and RAG on a single GPU.
Things to consider
- Supported languages are English, German, Spanish, French, Italian and Japanese.
- Reasoning traces lengthen responses; it can be switched to non-reasoning mode.
- Download sizes refer to the default quantized versions in the Ollama library; full-precision weights are larger.
Specifications are taken from the model card on Hugging Face, the Ollama library and the provider’s announcement. Benchmark results depend on evaluation settings; see how AI benchmarks work.