NVIDIA has added four open-weight checkpoints to its Cosmos 3 model family, aimed at generating interactive world simulations for autonomous driving and robotics. All four are post-trained variants of Cosmos 3 Nano, each with 16 billion trainable parameters, and are released under the OpenMDW 1.1 licence, which permits commercial and non-commercial use under its terms.
The driving models cover offline and streaming use. Cosmos3-Nano-Transfer-Auto generates synchronised video across up to 11 camera views, including fisheye lenses, and can jointly produce LiDAR data, letting developers vary weather, lighting and time of day to widen training and evaluation coverage. Cosmos3-Nano-Sim-Auto is a causally distilled streaming variant that produces short chunks of camera and LiDAR output in a few denoising steps, but closed-loop interaction still requires a separate simulation runtime and renderer. On the robotics side, Cosmos3-Nano-Sim-Transfer converts a physics simulator's depth, segmentation or edge controls into photorealistic video that a policy can consume, while Cosmos3-Nano-Sim-Bimanual predicts how a scene changes in response to two-arm actions for YAM and AgiBot embodiments. NVIDIA says physics and ground-truth state stay with the simulator, and it acknowledges that measuring how faithfully these models support policy learning and evaluation remains an open problem.
For Australian businesses and professionals, the release changes little unless they run robotics, autonomous-vehicle or simulation research teams. Running 16-billion-parameter models requires substantial GPU capacity, so the real cost is compute and engineering time rather than licensing. The licence allows commercial use, but any team that post-trains on real driving footage or recordings of people should check its obligations under the Privacy Act 1988, which the article does not address. For most other businesses there is nothing to act on yet; for those in the field, the sensible step is to benchmark the Transfer-Auto checkpoint against their own data before committing.
Key Points
NVIDIA added four open-weight Cosmos 3 Nano checkpoints, each with 16B trainable parameters, under the OpenMDW 1.1 licence
Transfer-Auto generates multi-view driving video across up to 11 cameras and can jointly produce LiDAR, with varied weather and lighting
Sim-Auto is a streaming variant that generates short camera and LiDAR chunks in a few denoising steps
Sim-Transfer turns simulator depth, segmentation or edge controls into video, and Sim-Bimanual predicts scene changes from robot actions
For Australian teams, the main cost is GPU compute, and any real-world training data raises Privacy Act questions