NVIDIA has launched the Cosmos 3 world model, making two versions available for download: Super and Nano. The Super version, with 64.6 billion parameters, is designed for applications requiring high physical accuracy, such as post-training robotics and autonomous driving. The Nano version, featuring 15.7 billion parameters, is optimized for low-latency scenarios like high-quality video and action reasoning. Both versions are accessible on Hugging Face and build.nvidia.com, supporting deployment as NVIDIA NIM microservices. Cosmos 3 is a multimodal foundational world model for physical AI, utilizing a Mixture of Transformers architecture to understand and generate text, images, video, environmental sounds, and actions. NVIDIA describes it as the first fully open multimodal model, allowing developers to download, fine-tune, and convert it into proprietary models. An Edge version for real-time inference is expected soon.