Cosmos 3 puts open world models behind robots, vehicles, and edge AI
Original: Into the Omniverse: How Open World Models Push the Frontier of Physical AI View original →
NVIDIA’s latest physical AI push is about collapsing several model jobs into one open family. In an August 6, 2026 post, the company framed Cosmos 3 as an open physical AI foundation omni-model family that combines vision reasoning, world generation, and action prediction. The target is practical: robots, autonomous vehicles, and vision AI systems need to understand scenes, generate synthetic data, predict future states, and train policies without every team maintaining a separate model stack.
The lineup spans three deployment profiles. Cosmos 3 Super is a 64B model for high-fidelity world modeling. Cosmos 3 Nano is a 16B model aimed at efficient reasoning and post-training. Cosmos 3 Edge is a 4B model built for on-device vision reasoning and robot policy deployment. NVIDIA says the Edge version can run across RTX GPUs, DGX systems, Jetson, and Jetson Thor platforms, which matters because physical AI often has to move from cloud simulation into local machines.
The benchmark claims are broad. NVIDIA says Cosmos 3 ranks No. 1 on Artificial Analysis for open weights text-to-image and image-to-video generation, on PAI-Bench for world generation, in the image-to-video category of Physics-IQ, and on RoboLab for robot policy. It also says Cosmos 3 Super is the highest-ranked open model on VANTAGE-Bench for vision understanding. The claim is not only image generation quality; it is a bid for a common foundation across perception, simulation, and action.
NVIDIA also named early users across industries: Doosan Robotics, LG Electronics, Samsung Electronics, and Skild AI in robotics; Li Auto, Xiaomi, and Afari in autonomous vehicles; plus several companies building industrial and smart-space vision agents. The model collection and datasets are available through Hugging Face and GitHub. The next proof point is deployment quality: whether these world models reduce the cost and risk of training physical systems once they leave benchmark suites and enter factories, roads, and edge devices.
Related Articles
NVIDIA released Cosmos 3 as an open physical AI omnimodel with Super and Nano variants. Its technical post points to six synthetic datasets, Hugging Face checkpoints, and GitHub recipes for domain adaptation.
NVIDIA and Hugging Face are bringing Isaac GR00T 1.7, Isaac Teleop, datasets, and robotics workflows into LeRobot. The move connects NVIDIA’s 3 million robotics developers with Hugging Face’s 16 million AI builders through a shared open robotics stack.
NVIDIA’s new Jetson T3000 and T2000 modules push Blackwell-class edge AI into smaller robotics systems. The real shift is cost and deployment: 865 FP4 TFLOPS, Cosmos 3 Edge, and memory-saving agent skills are being packaged for robots that cannot depend on cloud inference.