NVIDIA published four new open-weight checkpoints for the Cosmos 3 world model on Hugging Face. The release includes variants of Cosmos 3 Nano with 16B trainable parameters. The models use a Mixture-of-Transformers backbone for video prediction.
The Cosmos 3 Nano variants are Cosmos3-Nano-Transfer-Auto and Cosmos3-Nano-Sim-Auto. The Transfer variant generates synchronized video and LiDAR across up to 11 camera views. The Sim-Auto variant is a distilled model for fast streaming simulation with few denoising steps.
The release also provides post-training recipes for adapting the models to robotics and driving domains. Sim-Transfer generates visual observations from a physics simulator’s spatial controls. Sim-Bimanual simulates environments for bimanual robots like YAM and AgiBot. Developers can use these models for commercial and non-commercial purposes under the OpenMDW1.1 license. This gives researchers and engineers a new tool for training robotics and autonomous driving systems with open-source video models.
