What was released
NVIDIA released Cosmos 3 Edge on Hugging Face on July 20, 2026. The 4B world model is designed for robots and vision agents to understand scenes, predict future states, and connect actions with their visual consequences on edge hardware. The weights are downloadable without sign-in under the OpenMDW 1.1 license.
Model structure
- An autoregressive tower reads vision and text for understanding and reasoning.
- A diffusion tower handles prediction and generation over video, audio, and action tokens.
- Shared multimodal-attention layers connect scenes, possible futures, and actions.
- A DROID-post-trained manipulation policy and training scripts are also available.
NVIDIA reports that on Jetson Thor the model accepts 640×360 observations, generates 32 actions per inference, and supports 15 Hz control. It also released four-step distilled Cosmos3-Super image-to-video and text-to-image checkpoints, claiming up to 25 times faster inference than the base model.
Why it matters
It offers a path to move world modeling closer to factories, warehouses, and robots instead of relying only on a data-center API. “Open” does not make hardware or electricity free, and adopters must review NVIDIA-focused optimization and the OpenMDW terms.
Safety and validation
NVIDIA's safety file says the model is not safety certified and must not be the sole basis for life-critical decisions or control. Real robots need physical constraints, emergency stops, staged simulation and field tests, human supervision, and an independent safety controller. Rankings and speed claims should be rerun under the target hardware and workload.