NVIDIA releases Nemotron 3 Nano with open models and three trillion training tokens

NVIDIA/Nemotron 3

NVIDIA released Nemotron 3 Nano, a 30-billion-parameter open-weight model using hybrid mixture-of-experts to activate only 3 billion parameters per token. NVIDIA also published three trillion tokens of training and safety data for enterprises to fine-tune on proprietary data. Super and Ultra variants are planned for H1 2026.

• Nano immediately available on Hugging Face; Super (~100B active ~10B) and Ultra (~500B active ~50B) coming in first half of 2026.
• Hybrid latent MoE reduces inference compute while maintaining reasoning ability across multi-agent scenarios.
• Open datasets let enterprises customize without sharing proprietary data with NVIDIA.
• Trains efficiently on Blackwell GPUs using NVIDIA's 4-bit NVFP4 format.