NVIDIA Introduces Nemotron 3 Open Models for Multi-Agent AI Systems

NVIDIA has announced the Nemotron 3 family of open models, datasets, and libraries for building efficient and transparent multi-agent AI systems, announced in a press release. The lineup includes the Nemotron 3 Nano, Super, and Ultra models, each designed for scalability and efficiency using a hybrid latent mixture-of-experts architecture.

Nemotron 3 Nano, a 30-billion-parameter model activating up to 3 billion at a time, is available immediately. It delivers up to four times higher token throughput compared to Nemotron 2 Nano and reduces reasoning token generation by 60%, improving cost efficiency for applications such as debugging, summarization, and AI assistants. The model features a one-million-token context window, enhancing accuracy on long, multistep tasks.

Nemotron 3 Super and Ultra, with approximately 100 billion and 500 billion parameters respectively, are designed for more complex reasoning and multi-agent coordination. Both models use NVIDIA’s 4-bit NVFP4 training format on the Blackwell architecture to reduce memory use and speed up training. They are expected to become available in the first half of 2026.

Alongside the models, NVIDIA released open datasets and reinforcement learning tools including NeMo Gym, NeMo RL, and NeMo Evaluator. These resources are available on GitHub and Hugging Face to support model training, customization, and evaluation. Nemotron 3 Nano can be accessed through Hugging Face and inference providers such as Baseten, DeepInfra, Fireworks, FriendliAI, OpenRouter, and Together AI, as well as via AWS through Amazon Bedrock and other cloud platforms soon.