Nvidia launches open AI models Nemotron 3 Nano, Super, and Ultra to advance agentic AI, with Nano now available and Super/Ultra coming in 2026.

Nvidia has launched the Nemotron 3 family of open AI models—Nano, Super, and Ultra—designed to power advanced agentic AI systems. The Nano model, now available, features 30 billion parameters with 3 billion active, offering four times the throughput of its predecessor. Super and Ultra, with 100 billion and 500 billion parameters respectively, are set for release in early 2026. Built with a hybrid latent mixture-of-experts architecture, the models optimize memory and compute efficiency, enabling high performance on Nvidia’s Blackwell GPUs. Trained using a 4-bit format, they maintain accuracy while reducing memory use. The release includes 3 trillion tokens of training data, open-source reinforcement learning tools, and support for reasoning, tool use, and multi-agent collaboration. Early adopters include Accenture, Oracle, Palantir, and Zoom. While compatible with non-Nvidia hardware, the models are optimized for Nvidia platforms. The initiative aims to accelerate AI agent development, lower costs, and foster innovation across industries.