On December 15, 2025, NVIDIA unveiled its Nemotron 3 family of open models, aiming to revolutionize the development of agentic AI applications. Available in Nano, Super, and Ultra sizes, these models cater to increasing industry demands for efficient, high-performing multi-agent systems.
The Nemotron 3 models use a hybrid mixture-of-experts (MoE) architecture, promising exceptional accuracy and throughput. This launch places NVIDIA at the forefront of the open model sector, as businesses shift from traditional chatbots. The Nemotron 3 Nano is available immediately, boasting a fourfold throughput increase over its predecessor, the Nemotron 2 Nano, and is specifically targeted for software debugging and AI assistant tasks.
The Nemotron 3 family features three model versions. The Nano model includes 30 billion parameters, activating up to 3 billion for task management. The Super and Ultra models, with approximately 100 billion and 500 billion parameters respectively, are set to release in early 2026. NVIDIA claims these models will ease the transition to multi-agent environments, tackling communication overhead and context drift challenges.
Current partners like Accenture, Oracle, and Deloitte are already integrating these models. Despite unveiling the Nano model, NVIDIA has withheld pricing details for the Super and Ultra versions, presenting potential accessibility issues for smaller entities. Broader availability exists across platforms such as Hugging Face and major cloud services such as AWS and Google Cloud.
The efficiency and accuracy of Nemotron 3 remain unverified by independent benchmarks, making it challenging for potential users to gauge its merits against rivals like Meta’s Llama series. As companies transition to open models, transparency and coordination among multiple AI agents remain concerns. Upcoming details on the Super and Ultra models and further industrial applications will be pivotal for NVIDIA as it endeavors to validate its claims and enhance agentic AI capabilities across industries.
