NVIDIA has officially launched the Nemotron 3 series models, the most efficient family of open-source models to date, specifically designed for agentic AI applications. The series includes three models: Nano, Super, and Ultra, each excelling in cost efficiency, collaborative agents, and high-precision inference respectively. Nemotron 3 utilizes innovative technologies, including a hybrid Mamba-Transformer MoE architecture, latent expert design, and multi-token prediction, supporting context lengths up to 1 million tokens. The Nano model achieves inference efficiency 3.3 times greater than comparable models while maintaining high accuracy. NVIDIA has open-sourced the Nano model’s weights, training recipes, and related data, providing valuable resources to the AI community. This launch marks a significant breakthrough for NVIDIA in the field of efficient AI models and will have a profound impact on AI application development and deployment.
