NVIDIA Cosmos 3 Brings Open World Models to Robots, Vehicles and Edge AI
NVIDIA details Cosmos 3, an open world-model family with 64B, 16B and 4B variants for robotics, autonomous vehicles, simulation and edge AI at global scale.
Summary
Source-reported facts: In a blog post published Aug. 6, 2026, NVIDIA presented Cosmos 3 as an open family of foundation models for “physical AI” systems such as robots, autonomous vehicles and computer-vision agents. These world models are intended to learn relationships within physical environments, predict plausible future states and actions, and generate synthetic training data. NVIDIA argues that this approach can help developers address a central constraint in physical AI: collecting sufficient real-world data is costly, while dangerous, rare or unusual scenarios can be difficult to reproduce safely and repeatedly.
Cosmos 3 uses what NVIDIA describes as a mixture-of-transformers architecture combining visual reasoning, world generation and action prediction. The family has three principal sizes: the 64-billion-parameter Cosmos 3 Super for high-fidelity world modeling, the 16-billion-parameter Cosmos 3 Nano for more efficient reasoning and post-training, and the 4-billion-parameter Cosmos 3 Edge for on-device visual reasoning and robot-policy deployment. NVIDIA says Edge can run across RTX GPUs, DGX systems and Jetson hardware, including Jetson Thor. Developers can use the wider family as a vision-language model, a physics-grounded simulator, a synthetic-data generator or a foundation for specialized world-action models.
NVIDIA reports that Cosmos 3 leads several evaluations: Artificial Analysis for open-weights text-to-image and image-to-video generation, PAI-Bench for world generation, the image-to-video portion of Physics-IQ, and RoboLab for robot policy. It also says Cosmos 3 Super is the highest-ranked open model on VANTAGE-Bench for visual understanding. These results are claims in NVIDIA’s own announcement; the post does not provide enough methodology, competitor detail or external analysis to independently assess their broader significance.
The release sits within NVIDIA’s larger physical AI software strategy. Cosmos models are offered under the Linux Foundation’s OpenMDW 1.1 license, which NVIDIA says permits organizations to post-train them using their own data and hardware. Omniverse libraries provide tools for constructing simulation-ready environments, while OpenUSD supplies a framework for composing and exchanging complex 3D assets across digital twins, simulations and synthetic-data workflows. Related NVIDIA platforms include Isaac GR00T for robotics, Alpamayo for autonomous vehicles and Metropolis for vision AI. In July 2026, NVIDIA also joined more than 200 organizations in signing an open letter supporting open-weight models and a broad US AI ecosystem.
The company names Doosan Robotics, LG Electronics, Samsung Electronics and Skild AI among robotics developers using Cosmos; Li Auto, Xiaomi and Afari in autonomous vehicles; and Centific, Fogsphere, Linker Vision, Milestone Systems and Yuan in vision AI. NVIDIA is also expanding its Cosmos Coalition to Japan, where prospective participants plan work spanning manufacturing, logistics, agriculture, construction, healthcare and transportation. Interpretation and outlook: An adaptable model family could reduce the need to maintain separate systems for perception, generation and action prediction, but the announcement does not quantify production deployments, cost savings, safety improvements or real-world reliability. Developers can next examine the models and datasets on Hugging Face and GitHub, review NVIDIA’s technical report and conduct their own tests. Whether Cosmos 3’s benchmark strength translates into dependable operation under unusual physical conditions remains uncertain.
Positives
- NVIDIA offers Cosmos 3 under the Linux Foundation’s OpenMDW 1.1 license, allowing teams to post-train the models on their own data and hardware.
- The family spans three deployment profiles: 64B Cosmos 3 Super, 16B Cosmos 3 Nano and 4B Cosmos 3 Edge, giving developers options from high-fidelity modeling to on-device inference.
- NVIDIA says Cosmos 3 ranks first on PAI-Bench for world generation and RoboLab for robot policy, while Cosmos 3 Super leads open models on VANTAGE-Bench for visual understanding.
- Cosmos 3 Edge is designed for deployment across RTX GPUs, DGX systems and Jetson devices, including Jetson Thor platforms.
- NVIDIA identifies developers across robotics, autonomous vehicles and vision AI, including Doosan Robotics, LG Electronics, Samsung Electronics, Li Auto, Xiaomi and Milestone Systems.
- The Cosmos Coalition is expanding to Japan, with planned work on open world models for sectors including factories, logistics, agriculture, healthcare and transportation.
Risks & concerns
- NVIDIA acknowledges that physical AI training data is expensive to collect and that rare or long-tail events are especially difficult to reproduce safely and consistently.
- A general world model will not initially understand a developer’s specific robot, sensors, task or operating environment, so additional data and post-training remain necessary.
- The benchmark leadership claims come from NVIDIA’s own blog post, which does not include enough methodological or comparative detail to establish how the results translate to real deployments.
- The article lists numerous companies building with Cosmos but does not quantify production-scale deployments, performance gains, safety outcomes or cost reductions.
- Combining world generation, visual reasoning and action prediction does not eliminate uncertainty about reliability when systems encounter unusual or previously unseen physical conditions.