Nvidia published four open-weight Cosmos 3 models on Hugging Face on November 5 for interactive world simulation
Nvidia published four open-weight checkpoints post-trained from Cosmos 3 Nano covering multi-view synthetic driving data generation, a causal distilled streaming driving simulator, simulator-to-photorealistic-video conversion and a bimanual next-frame predictor, all under OpenMDW-1.1.
On November 5, Nvidia published the Cosmos 3 open models on Hugging Face, comprising four checkpoints post-trained from Cosmos 3 Nano at roughly 15 to 16 billion parameters each, targeting multi-view synthetic driving data generation, causally distilled streaming driving simulation, simulator-controls-to-photorealistic-video conversion and bimanual robot next-frame prediction.
The multi-view synthetic driving data generation model can output up to 11 camera views, the streaming driving simulator achieves low-latency generation through causal distillation, the simulator-to-video model converts rendered scenes into photorealistic video, and the bimanual model predicts the next frame after a robot action. All four checkpoints are released under the OpenMDW-1.1 license allowing commercial and non-commercial use, with training recipes and agent skills also open-sourced on GitHub.
The models are positioned as a lower-cost entry point into simulation data and simulation environments for robotics and autonomous driving teams, so they can validate simulation pipelines with open weights before committing to more expensive commercial options. For embodied AI infrastructure teams this means the simulation chain can be validated first and the procurement decision deferred.