NVIDIA adds four open Cosmos 3 checkpoints for driving and robot simulation
The checkpoints generate multi-view driving data, synchronized camera and LiDAR streams, photorealistic video from simulator controls, and video predicted from robot actions. All four are post-trained versions of Cosmos 3 Nano under the OpenMDW 1.1 license.
Why it matters: Open world models let robotics and driving teams test against generated environments, though NVIDIA says how faithfully they support real policy learning is still being measured.
- Trainable parameters
- 16B each
- Weights size
- 31.7 GB to 34.3 GB each (Hugging Face)
- License
- OpenMDW 1.1
Sources: huggingface.co