According to Dynamic Beating monitoring, NVIDIA has officially released the model weights of Cosmos 3 World Model for download today. The first batch includes two versions: Super (646 billion parameters) and Nano (15.7 billion parameters), both of which are now available on HuggingFace (ungated for direct download) and build.nvidia.com. They also support deployment in the form of NVIDIA NIM microservices.
Cosmos 3 is positioned as a full-moda (omnimodel) world base model for physical AI, based on a new hybrid Transformer architecture (Mixture of Transformers), with native understanding and generation of text, image, video, environmental sound, and motion. The Super version targets post-training robots and autonomous driving models that require the highest physical accuracy, while the Nano version is designed for low-latency scenarios focusing on high-quality video and motion inference. In addition, an Edge version (aimed at edge-side real-time inference) is expected to be released soon.
NVIDIA claims that Cosmos 3 is the "world's first fully open full-moda model," allowing developers to freely download, fine-tune, and convert it into proprietary models.
