Nemotron 3: Supercharge Agentic AI Now

NVIDIA Expands Open Model Families to Power Agentic, Physical, and Healthcare AI

NVIDIA is revolutionizing AI development by releasing expansive families of open models, data, and tools tailored for agentic systems, physical AI applications, and healthcare innovations. These resources empower developers worldwide to build efficient, specialized AI solutions across industries, accelerating from prototypes to real-world deployment.

The Nemotron 3 Family: Fueling Agentic AI

NVIDIA's Nemotron 3 family stands at the forefront of agentic AI, offering open models designed for transparent and efficient development. These models leverage advanced reinforcement learning techniques and concurrent multi-environment post-training at scale to deliver superior accuracy in complex workflows.

Key variants include Nemotron 3 Super with 49 billion parameters and Nemotron 3 Ultra with 253 billion parameters, both slated for release in the first half of 2026. Trained using NVIDIA's ultra-efficient 4-bit NVFP4 format on the Blackwell architecture, they reduce memory needs and speed up training without sacrificing performance.

• Enable scaling from dozens to hundreds of agents for long-horizon reasoning.
• Support startups in building AI teammates for human-AI collaboration.
• Form part of the NVIDIA Nemotron Coalition, uniting AI developers for collaborative advancement.

Advancing Physical AI with Cosmos and Specialized Models

For physical AI and robotics, NVIDIA introduces the Cosmos platform, featuring open world foundation models that provide human-like reasoning and world generation. These tools accelerate development and validation in complex real-world environments.

Built on Cosmos are targeted models such as:

• Isaac GR00T N1.6, a vision-language-action model for humanoid robots enabling full-body control and enhanced contextual understanding.
• Alpamayo family for autonomous vehicle development, supporting reasoning in driving scenarios.
• Blueprints for video search and summarization within the Metropolis platform to boost operational efficiency.

NVIDIA also shares vast open datasets, including 500,000 robotics trajectories and 100 terabytes of vehicle sensor data, fostering innovation in embodied intelligence.

Transforming Healthcare and Beyond with Clara and More

In healthcare, the Clara family of open models targets biomedical applications, complemented by 455,000 protein structures in open datasets. This expands to other domains like climate science with Earth-2 and multimodal AI via Nemotron.

NVIDIA's strategy emphasizes openness: models trained on supercomputers top leaderboards and are freely available for creation, evaluation, guardrailing, and deployment. The portfolio spans six key areas, creating a global ecosystem for every industry and country.

A New Era of Open AI Innovation

NVIDIA's comprehensive release of open models, training frameworks, and multimodal data at unprecedented scales marks a pivotal shift. From agentic reasoning to robotics and healthcare, these tools lower barriers, cut costs, and drive breakthroughs on platforms like Rubin and Blackwell.

Developers gain right-sized models for diverse workloads, ensuring faster iteration and enterprise-scale deployment. This open approach positions NVIDIA as a leader in democratizing frontier AI.