NVIDIA published four new Cosmos 3 Nano open-weight checkpoints in its November 5 technical release, aiming to make interactive world simulation more accessible to physical-AI researchers. The systems address different inputs and outputs rather than acting as interchangeable general-purpose language models. NVIDIA identifies each as a post-trained example of Cosmos 3 Nano with 16 billion trainable parameters. Their practical value lies in generating visual observations conditioned on traffic scenes, simulator states or robot actions, while the accuracy of those simulated consequences still needs to be established for each application. Primary source ↗
Cosmos3-Nano-Transfer-Auto is designed for offline synthetic driving data and can generate synchronized views across up to eleven cameras, including fisheye configurations, with joint RGB and LiDAR options. Cosmos3-Nano-Sim-Auto moves toward sequential interaction by streaming observations as the driving-control inputs change. Researchers could use these outputs to explore robustness across weather, lighting and traffic conditions, but a visually plausible rendered road scene is not evidence of physically correct vehicle behavior.
The remaining two variants address different forms of controllability. Cosmos3-Nano-Sim-Transfer turns simulator controls and text conditioning into rendered RGB observations, supporting workflows built around an existing physics simulation. Cosmos3-Nano-Sim-Bimanual predicts scenes following two-arm robot or camera actions, with a specific action representation documented by NVIDIA. Combining camera motion, action timing and observed state is a useful research direction, although error accumulation over long sequences and real-world policy transfer are major unresolved evaluation questions.
NVIDIA lists the releases under OpenMDW 1.1, which permits commercial and non-commercial activity subject to license conditions rather than carrying the unconditional implications sometimes associated with permissive software licenses. The published deployment recipe uses Linux, NVIDIA GPU infrastructure and supporting containers; a demonstrated GB200 environment should not be taken as a universal minimum specification. Open Model Weights treats this as four identifiable source-published checkpoint variants, not proof of real-robot performance or independent world-model fidelity.
What this changes for the evidence layer
These four Cosmos3-Nano post-trained checkpoints are not a general claim that simulated actions transfer faithfully to hardware. Verify each exact repository revision, OpenMDW 1.1 licensing terms and runtime dependencies before any curated model update or hardware deployment.
NVIDIA — Cosmos 3 interactive world simulation technical blog
This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.
Open source article ↗