Cosmos 3: A Unified Omnimodal World Model for Advancing Physical AI
Cosmos 3 is a novel omnimodal world model built on a mixture-of-transformers architecture, designed to unify and process diverse data modalities including language, image, video, audio, and actions. This unified framework enables state-of-the-art performance across various AI tasks, positioning it as a versatile backbone for embodied agents and advancing the field of Physical AI. The project emphasizes open research by releasing code, model checkpoints, datasets, and benchmarks under the OpenMDW-1.1 License.