NVIDIA Open-Weights Four Cosmos 3 Checkpoints for Interactive World Simulation

📅 November 5, 2026 • NVIDIA

NVIDIA has released four open-weight Cosmos 3 Nano checkpoints for interactive world simulation, per the NVIDIA blog on Hugging Face on November 5, 2026. The checkpoints target two domains where world models are becoming working infrastructure: driving and robotics. The weights are hosted on Hugging Face, and alongside them NVIDIA published post-training recipes and coding-agent skills — material meant to help teams adapt the models to their own sensor rigs and robot embodiments.

Two named variants cover complementary tasks. Transfer-Auto turns a scenario into synchronized multi-sensor data, generating up to 11 camera views — including fisheye — with joint RGB-LiDAR generation, so camera and LiDAR outputs stay consistent within one pass. That is the data-generation case: synthetic multi-sensor driving data from a scenario description. Sim-Auto is the interactive case: it streams camera and LiDAR observations chunk by chunk from controls, so a driving policy can act, receive fresh observations, and act again inside a rolling simulation. The other two checkpoints round out the family with further driving and robotics simulation variants.

The architecture is documented as well. The Cosmos 3 backbone is a Mixture-of-Transformers, a design that links text and visual networks rather than folding both modalities into one monolithic model. NVIDIA positions the release as an open foundation for physical AI: open weights, post-training recipes, and coding-agent skills that help assemble a working pipeline.

For context, world models of this kind sit directly on the path of autonomous-vehicle and robotics development, where real sensor data is expensive to collect and edge cases are what matters most. Teams can vary conditions in simulation, generate matched multi-sensor ground truth, and evaluate policies before anything touches a real vehicle or robot. Open weights change the economics of that loop, because the generator becomes something a team can inspect, post-train and version-control. Outside NVIDIA, the release has so far been picked up by Gate News and explainx.ai, which cite the same NVIDIA materials.

Impact: for AV and robotics teams, the value is an openly available path to synthetic multi-sensor data and interactive simulation. A perception team can generate synchronized camera-plus-LiDAR scenario variations through Transfer-Auto, and a policy team can evaluate against Sim-Auto's streamed observations, adapting both to a custom sensor setup with the published recipes. Open checkpoints also make results comparable across teams, since everyone evaluates against the same artifacts rather than private renders, and they lower the entry barrier for labs that cannot license a closed simulation stack.

Sources:
https://huggingface.co/blog/nvidia/cosmos-interactive-world-simulation
https://developer.nvidia.com/blog/build-world-models-for-interactive-simulation-with-nvidia-cosmos
https://explainx.ai/blog/nvidia-cosmos-interactive-simulation-guide-sim-auto-transfer-auto-h-dreams-nim-vram-table-2026