NVIDIA Unveils Nemotron 3: A New Era for Open AI Models

NVIDIA has announced the launch of its Nemotron 3 family of open models, marking a significant advancement in the development of efficient and specialized agentic AI applications. The new series offers three model sizes—Nano, Super, and Ultra—each designed to cater to different levels of complexity and scale in AI systems.

Innovative Architecture for Enhanced Performance

The Nemotron 3 models introduce a groundbreaking hybrid latent mixture-of-experts (MoE) architecture. This new design is aimed at overcoming the challenges faced by developers when transitioning from single-model chatbots to multi-agent AI systems, such as communication overhead and high inference costs. According to NVIDIA, the Nano variant of Nemotron 3 delivers four times higher throughput than its predecessor, optimizing performance for multi-agent systems at scale.

Revolutionizing AI Development

Jensen Huang, founder and CEO of NVIDIA, emphasized the importance of open innovation in AI, stating that Nemotron 3 transforms advanced AI into an open platform that offers the transparency and efficiency necessary for building agentic systems. The models are part of NVIDIA’s broader efforts to support sovereign AI initiatives globally, with organizations in Europe and South Korea adopting these open models to align AI systems with regional data and regulatory standards.

Industry Adoption and Impact

Early adopters such as Accenture, Deloitte, and Oracle Cloud Infrastructure are integrating Nemotron models into their workflows across various industries, including manufacturing and cybersecurity. This adoption underscores the models' versatility and effectiveness in enhancing AI-driven processes. Bill McDermott, CEO of ServiceNow, highlighted the collaboration with NVIDIA as a significant step forward in empowering industry leaders to accelerate their AI strategies.

Technical Specifications and Availability

The Nemotron 3 Nano, optimized for tasks such as software debugging and content summarization, is currently available and is recognized for its cost-efficiency and scalability. The Super and Ultra models, expected in 2026, will cater to more complex applications requiring advanced reasoning capabilities. These models employ NVIDIA’s ultraefficient 4-bit NVFP4 training format, which reduces memory requirements and accelerates training without sacrificing accuracy.

Supporting Tools and Resources

NVIDIA has also released comprehensive training datasets and reinforcement learning libraries to aid developers in building specialized AI agents. These resources, available on platforms such as GitHub and Hugging Face, provide the necessary tools for creating domain-specialized AI agents with enhanced safety and performance.

This launch positions NVIDIA at the forefront of AI innovation, offering developers a robust platform to build cutting-edge AI solutions tailored to modern industry needs.

For further information, visit the official NVIDIA Newsroom.