NVIDIA releases four open-weight Cosmos 3 Nano checkpoints for interactive driving and robot world simulation

NVIDIA put four post-trained, 16-billion-parameter versions of its Cosmos 3 Nano model on Hugging Face. They cover multi-view driving video (up to 11 cameras plus LiDAR), streaming camera and LiDAR simulation, simulator controls plus text to RGB video, and action-conditioned video for YAM and AgiBot robots. The weights are under the OpenMDW 1.1 licence, which allows commercial and non-commercial use. A technical report, notebooks and inference "skills" came with them. The extraction found no primary evidence links, so the details come from NVIDIA's own post and are not independently checked.

The main audience is robotics and autonomous-driving teams that want a free, commercially usable starting point for simulated training environments. They can now test whether these models fit their own data and sensor setups without building from scratch. The impact is limited by the niche: these are fine-tuned variants of an existing model family, and the release is useful mainly to teams already working with these sensor types or robot platforms. Competitors in simulation tooling get another free option to compare against.

Phrases like "shared starting point" and "interactive simulated worlds" are promotional framing. The release gives no benchmark evidence that the models perform well, and NVIDIA itself says their limits still need to be measured. Open weights also do not mean fully open: the licence terms, training data and real-world fidelity need checking before anyone relies on the models.

1. 01Independent evaluations of the checkpoints on driving and bimanual-robot tasks, such as how well simulated video matches real sensor data, within the next one to two months
2. 02Hugging Face download counts and community fine-tunes or derived projects, which would show whether developers actually adopt these over existing simulators
3. 03Close reading of the OpenMDW 1.1 licence and the technical report for restrictions, hardware requirements (GPU memory for 16B models) and any disclosed training-data limits