NVIDIA has introduced Cosmos 3, an open-world model family designed for physical AI applications. The model combines vision reasoning, world generation, and action prediction to help developers understand scenes, generate synthetic data, and simulate future states. Developers can use Cosmos 3 as a vision language model, a physics-grounded world simulator, or the backbone for world action models, eliminating the need for separate models for each capability. The family includes three variants tailored for different use cases: Cosmos 3 Super (64B), Cosmos 3 Nano (16B), and Cosmos 3 Edge (4B). Cosmos 3 Edge is lightweight enough to run on edge GPUs, including NVIDIA RTX GPUs, DGX systems, and Jetson platforms, making it suitable for on-device vision reasoning and robot policy deployment. Across benchmark evaluations, Cosmos 3 ranks No. 1 on Artificial Analysis for open weights text-to-image and image-to-video generation, on PAI-Bench for world generation, and in the image-to-video category of Physics-IQ. For robot policy, it ranks No. 1 on RoboLab. Cosmos 3 Super is also the highest-ranked open model on VANTAGE-Bench for vision understanding. The release of Cosmos 3 is part of NVIDIA’s broader physical AI stack, which includes Isaac GR00T for robotics, Alpamayo for autonomous vehicles, and Metropolis for vision AI. The NVIDIA Cosmos Coalition has also expanded to Japan, where robotics and manufacturing leaders plan to collaborate on open world models for various industries. Developers from companies like Doosan Robotics, LG Electronics, and Li Auto are already using Cosmos 3 for physical AI applications. The model’s open licensing under the Linux Foundation’s OpenMDW 1.1 license allows teams to post-train models on their own data and hardware. This openness is essential for specialization, as it enables teams to adapt models to specific robots, vehicles, or operating environments. The combination of Cosmos 3 with NVIDIA Omniverse libraries and OpenUSD provides developers with the tools needed to build simulation-ready environments for training, testing, and validating physical AI systems before real-world deployment. The release of Cosmos 3 marks a significant step in the development of open-world models for physical AI, offering a flexible and adaptable foundation for robotics, autonomous vehicles, and vision AI systems.

NVIDIA’s approach to physical AI emphasizes the importance of open models that can be customized for specific applications. World models are critical for physical AI because they must understand and predict consequences, not just appearances. These models learn how physical environments behave, what may happen next, and which actions make sense. By generating physically grounded world and action data, they provide a foundation that teams can specialize for robots, autonomous vehicles, or vision AI systems. Open world models are already being used to generate training data, test policies, and specialize physical AI systems. The open nature of these models allows for post-training and adaptation, which is essential for addressing the unique requirements of different applications. The NVIDIA Cosmos Coalition, which includes world model builders, AI developers, and physical AI leaders, is working to contribute models, research, and evaluation methods to advance the field. This collaboration is helping to establish open world models as an adaptable foundation for physical AI across various domains. The release of Cosmos 3 and the expansion of the NVIDIA Cosmos Coalition highlight the growing importance of open-world models in the development of physical AI systems.

NVIDIA has emphasized the need for open ecosystems in AI development, as stated in the Open Weights and American AI Leadership open letter. This initiative argues that AI leadership should be measured by the reach of an open ecosystem across all sectors. Open models, which can be downloaded, inspected, modified, and run on any infrastructure, are essential for this goal. The release of Cosmos 3 aligns with this vision, as it provides developers with the tools and flexibility needed to build and adapt physical AI systems. The open licensing of Cosmos 3 under the Linux Foundation’s OpenMDW 1.1 license enables teams to post-train models on their own data and hardware, fostering innovation and customization. This approach is crucial for addressing the specialized needs of different applications, from robotics to autonomous vehicles. The combination of Cosmos 3 with NVIDIA Omniverse libraries and OpenUSD further enhances the capabilities of physical AI teams by providing the tools needed to build simulation-ready environments. This integration reduces duplicated work and streamlines the development process, allowing teams to focus on creating effective solutions for real-world applications. The release of Cosmos 3 and the expansion of the NVIDIA Cosmos Coalition represent significant progress in the development of open-world models for physical AI, offering a flexible and adaptable foundation for various applications.

Source: nvidia