The useful question is what changes for users, developers or buyers, and whether the announcement stays industry context or becomes something people can actually use.

NVIDIA has unveiled Cosmos 3, a new family of open world foundation models designed to advance physical AI by predicting environmental behaviors and generating synthetic training data. The models support specialization for robots, autonomous vehicles, and vision systems, addressing challenges in data collection and rare event simulation. Cosmos 3 is available under the Linux Foundation’s OpenMDW 1.1 license, allowing teams to adapt models to their specific hardware and operating conditions.

The Cosmos 3 family includes three variants: Cosmos 3 Super (64B) for high-fidelity modeling, Cosmos 3 Nano (16B) for efficient reasoning, and Cosmos 3 Edge (4B) for on-device deployment. These models integrate vision reasoning, world generation, and action prediction, reducing the need for separate specialized models. Benchmarks show Cosmos 3 leading in text-to-image, image-to-video, and world generation tasks, with top rankings in Physics-IQ, RoboLab, and VANTAGE-Bench.

NVIDIA’s physical AI stack includes additional tools like Isaac GR00T for robotics, Alpamayo for autonomous vehicles, and Metropolis for vision AI. Developers across industries, including Doosan Robotics, LG Electronics, and Samsung Electronics, are already leveraging Cosmos for applications in robotics, autonomous driving, and industrial AI systems.

To foster collaboration, NVIDIA has launched the Cosmos Coalition, expanding its reach to Japan to develop open world models for manufacturing, logistics, and healthcare. The models and datasets are accessible on Hugging Face and GitHub, supporting broader adoption and innovation in physical AI.