NVIDIA offers downloadable Cosmos 3 checkpoints, inference skills, and notebooks

Top News

NVIDIA's Cosmos 3 is a world-model development platform for robotics and autonomous-driving simulation. It provides four downloadable open-weight checkpoints, with inference skills and notebooks to run them.

Here's what you can build with it:

• Transfer-Auto generates driving data with up to 11 synchronized camera views and joint RGB–LiDAR output.
• Sim-Auto streams synchronized camera and LiDAR observations, using generated history and updated driving controls.
• Sim-Transfer renders video from simulator controls and text, while an external simulator handles motion, contacts, and physical state.
• Sim-Bimanual predicts video from robot or camera actions. Its documented setup requires Linux, an NVIDIA GPU, Docker, and NVIDIA Container Toolkit.
• All four checkpoints have 16B trainable parameters and use OpenMDW1.1, which permits commercial and non-commercial use under its terms.

One catch: How faithfully these models support policy learning and evaluation remains an open question. A successful short training run does not establish convergence.

Sources: Hugging Face · arXiv paper