NVIDIA Nemotron

Nemotron is NVIDIA's umbrella for open models aimed at agentic AI, and it is unusual in shipping not just weights but the datasets and step-by-step training recipes behind them. The Nemotron 3 generation uses a hybrid Mamba-transformer mixture-of-experts backbone and spans Nano (31.6B total, 3.6B active) for edge and PC deployment, Nano Omni (30B/3B) for multimodal input, Super (120.6B/12.7B) for single-GPU throughput, and Ultra at 550B/55B for multi-GPU datacenters, with context windows reaching 1 million tokens; earlier Llama-Nemotron variants remain available. The GitHub repository collects cookbooks, deployment guides, and end-to-end examples under Apache-2.0, while model weights use the NVIDIA Open Model License, which permits commercial use, modification, redistribution, and hosted services without royalties. Pre-training, post-training, and reinforcement learning datasets are published openly on Hugging Face, and official documentation lives at docs.nvidia.com/nemotron. For teams standardizing on NVIDIA hardware the family offers a complete open pipeline from data to deployed agent.