Nvidia Debuts Nemotron 3 Family of Open Models (nvidianews.nvidia.com)

NVIDIA has announced the Nemotron 3 family of open models, which includes the Nano, Super, and Ultra sizes, designed to enhance the development of agentic AI applications. This new suite utilizes a breakthrough hybrid mixture-of-experts (MoE) architecture that significantly boosts model efficiency, with Nemotron 3 Nano achieving up to 4x higher throughput than its predecessor. The models leverage advanced reinforcement learning techniques and a large context window of 1 million tokens, improving accuracy in complex, multi-agent workflows while maintaining low inference costs. This launch is pivotal for the AI/ML community as it marks NVIDIA's commitment to open innovation, providing organizations with the tools to create specialized, transparent AI systems that accommodate their particular data and regulatory needs. Key technical advancements include the introduction of training datasets and libraries, with three trillion tokens of Nemotron pretraining data available to enhance agent performance. The models are expected to drive innovation from startups to enterprise solutions, enabling efficient collaboration among AI agents across various industries including manufacturing, cybersecurity, and software development. The Nemotron 3 family is now available via platforms like Hugging Face and is set to revolutionize how developers approach multi-agent AI system design.