New NVIDIA Cosmos 3 Open Models for Interactive World Simulation

NVIDIA has published Cosmos 3, a new generation of open models under its Cosmos world foundation model family, focused specifically on interactive world simulation. The release is hosted on Hugging Face and targets researchers and developers working on embodied AI, robotics, and autonomous vehicle development - fields where realistic, controllable simulated environments are increasingly important for training and evaluation.

World simulation models like Cosmos 3 are designed to generate video of physical environments that respond coherently to actions or inputs, rather than simply producing passive scene footage. This interactivity makes them particularly useful as synthetic data sources or as lightweight world models within reinforcement learning pipelines, where agents need to predict the consequences of their actions without running expensive real-world experiments.

Cosmos 3 builds on NVIDIA's earlier Cosmos releases, which established a foundation for video generation grounded in physical plausibility. The move to open weights continues a broader pattern in the generative AI space of large hardware and cloud companies releasing model weights to the research community, both to drive adoption of their platforms and to accelerate external research. Hosting on Hugging Face makes the models straightforwardly accessible to practitioners already working within that ecosystem.

Details on the specific architecture changes, context length, resolution capabilities, and benchmarked performance of Cosmos 3 relative to its predecessors are available in the full technical writeup on Hugging Face. For teams working on simulation-heavy workflows - particularly in autonomous systems - this release is worth evaluating as a potential component in synthetic data generation or world modeling pipelines.