NVIDIA details four open-weight Cosmos 3 checkpoints for driving and robot simulation
NVIDIA's 16B post-trained models cover multi-view driving data, synchronized camera-LiDAR streams, simulator-to-video rendering and two-arm robot prediction.
By Ryan Merket · Published
Primary source: Hugging Face Newsroom
Why it matters
Open, task-specific world-model checkpoints lower the cost of generating rare driving conditions and robot training scenes without collecting every case physically. The commercial-use license and adaptation recipes make the release usable inside company simulation stacks.
NVIDIA added four open-weight Cosmos 3 Nano checkpoints to Hugging Face on November 5th, 2026, extending the Cosmos 3 family into streaming driving simulation and action-conditioned robot video. The four releases are post-trained versions of Cosmos 3 Nano, each with 16B trainable parameters, released under OpenMDW 1.1 terms that permit commercial and non-commercial use.…
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.