NVIDIA introduces Nemotron 3 family of open models – Nano, Super, Ultra sizes for building specialized AI agents with superior accuracy and efficiency for multi-agent systems at scale. NVIDIA releases training datasets, reinforcement learning environments, and libraries for transparent and efficient AI development across industries.

Nemotron 3 Nano offers 4x higher throughput than Nemotron 2 Nano, with a breakthrough hybrid MoE architecture for multi-agent systems. NVIDIA’s open models are adopted by organizations worldwide for AI development aligned with data, regulations, and values. Early adopters integrate Nemotron models for AI workflows in manufacturing, cybersecurity, software development, and more.

Nemotron 3 models include Nano, Super, Ultra sizes, providing efficient reasoning engines for targeted tasks, multi-agent applications, and complex AI workflows. Nemotron 3 Nano delivers higher token throughput and accuracy, ranking as the most open and efficient model in its size category. Super and Ultra models excel in applications requiring collaborating agents and advanced reasoning.

NVIDIA’s Nemotron 3 models allow developers to choose the right-sized model for specific workloads, scaling from dozens to hundreds of agents for faster, more accurate reasoning in complex workflows. NVIDIA also releases training datasets, reinforcement learning libraries, and tools for building domain-specialized AI agents. Open-source NeMo Gym and NeMo RL libraries provide training environments for Nemotron models.

Nemotron 3 Nano is available for deployment on Hugging Face and inference service providers for efficient software debugging, content summarization, and AI assistant workflows. Nemotron 3 Super and Ultra models are expected to be available in the first half of 2026. NVIDIA supports Nemotron models on enterprise AI platforms and public clouds for secure, scalable deployment.

Read more at Nvidia News: NVIDIA Debuts Nemotron 3 Family of Open Models