NVIDIA Launches Cosmos 3 World Model with Two Versions Available for Download

According to Beating Monitor, NVIDIA has officially opened the download of model weights for Cosmos 3 today, initially releasing two versions: Super (64.6 billion parameters) and Nano (15.7 billion parameters), both now available on Hugging Face (open access, no gatekeeping) and build.nvidia.com, with support for deployment as NVIDIA NIM microservices. Cosmos 3 is positioned as a multimodal foundational world model for physical AI, built on a novel Mixture of Transformers architecture, natively understanding and generating text, images, video, environmental sounds, and actions. The Super version targets post-training robotics and autonomous driving models requiring the highest physical accuracy, while the Nano version is optimized for low-latency scenarios involving high-quality video and action reasoning. Additionally, an Edge version (designed for edge-based real-time inference) is expected to be released soon. NVIDIA describes Cosmos 3 as “the world’s first fully open multimodal model,” allowing developers to freely download, fine-tune, and convert it into proprietary models.