NVIDIA Cosmos 3 advances open world models for physical AI

In July, NVIDIA joined more than 200 companies and organizations in backing an open letter that tied AI leadership to broad, open ecosystems rather than a single frontier model. NVIDIA argues that openness is especially important for physical AI, where systems must understand consequences in the real world and adapt to specific robots, sensors, vehicles and operating environments.

World models help generate training data, simulate future states, test policies and specialize physical AI systems. NVIDIA Cosmos world foundation models are available under the Linux Foundation’s OpenMDW 1.1 license, allowing teams to post-train models on their own data and hardware. Omniverse libraries and OpenUSD support the workflow by helping developers build simulation-ready environments and reuse complex 3D data across digital twins, simulations and synthetic data generation.

NVIDIA Cosmos 3 combines vision reasoning, world generation and action prediction in an open physical AI foundation model family. The lineup includes Cosmos 3 Super (64B), Cosmos 3 Nano (16B) and Cosmos 3 Edge (4B), with the edge model designed for on-device vision reasoning and robot policy deployment. NVIDIA says Cosmos 3 ranks No. 1 across several open-weight generation and physical AI benchmarks, while companies including Doosan Robotics, LG Electronics, Samsung Electronics, Li Auto, Xiaomi and others are building on Cosmos for robotics, autonomous vehicles and vision AI agents.