NVIDIA Cosmos 3 is a new open physical artificial intelligence foundation model designed for physical reasoning and world simulation. The system combines vision reasoning, world generation, and action prediction. Consequently, developers can use the model to build physical AI applications with higher efficiency.
Features of NVIDIA Cosmos 3
The model operates as a fully open omnimodel. Specifically, it processes text, images, video, ambient sound, and actions. This capability helps reduce training cycles for physical AI from months to days. Meanwhile, the system delivers high physics accuracy across multiple benchmarks.
A New Architecture for Physical AI
The system uses a mixture-of-transformers architecture. This setup pairs a reasoning transformer with an expert generation transformer. As a result, the model understands object interactions and spatial-temporal relationships before generating video trajectories. This architecture helps robots and autonomous vehicles generalize in real-world environments.
NVIDIA trained the NVIDIA Cosmos 3 model on a large multimodal dataset. This dataset includes billions of samples across text, image, video, sound, and action trajectories. Therefore, developers can build physical AI systems with less data and lower training costs. The model ranks first on several open benchmarks, including Physics-IQ and R-Bench.
The Cosmos Coalition Collaboration
To support this technology, the company launched the Cosmos Coalition. This global group includes AI developers and robotics companies such as Agile Robots, Black Forest Labs, Runway, and Skild AI. Members will contribute models, research, and evaluation techniques. Furthermore, they will use the shared infrastructure for large-scale training.
Jensen Huang, founder and CEO of NVIDIA, stated that physical AI development is accelerating due to breakthroughs in multimodal reasoning. He noted that the new open models provide developers with improved capabilities to build robots and autonomous systems that perceive, reason, plan, and act in the physical world.
Model Availability and Deployment
The NVIDIA Cosmos 3 lineup includes different versions for various development stages. Specifically, the Super version is designed for post-training robotics, while the Nano version provides rapid video and action reasoning. The company plans to release an Edge version soon for real-time inference. Developers can access these models on Hugging Face and GitHub.
In addition, several technology companies are already building on the platform. These partners include Samsung, LG Electronics, and Li Auto. They use the tools for robotics and autonomous driving applications. Ultimately, the platform aims to simplify the deployment of physical AI across multiple industries.
