Nvidia Nemotron 3 Family of Models (research.nvidia.com)

NVIDIA has unveiled the Nemotron 3 family of models, marking a significant advancement in open models for agentic AI applications. Comprising three variants—Nano, Super, and Ultra—this lineup offers enhanced reasoning, conversational abilities, and efficiency. The Nano model, at 3.2 billion parameters, excels in accuracy while maintaining cost-effective inference speeds, outperforming notable competitors like GPT-OSS-20B. The Super and Ultra models are tailored for demanding tasks such as IT automation and are built on a novel hybrid Mamba-Transformer architecture, leveraging Latent MoE for improved accuracy and efficiency. The significance of Nemotron 3 lies in its advanced technical features, including support for up to 1 million tokens in context length and a multi-environment reinforcement learning post-training approach, ensuring superior adaptability across diverse tasks. Additionally, the emphasis on open sourcing the model weights, training recipes, and extensive datasets, which include over 2.5 trillion tokens, fosters collaboration and innovation in the AI/ML community. These models promise not only to enhance the performance of AI applications but also to facilitate research and development in open-source AI, positioning NVIDIA at the forefront of the agentic AI revolution.