NVIDIA expands open world models for physical AI development
Fri, 7th Aug 2026 (Today)
NVIDIA has outlined its push for open world models for physical AI, centred on its Cosmos 3 model family and Omniverse software tools.
Open models are becoming central to robotics, autonomous vehicles and vision AI, where systems must be adapted to specific machines, sensors and operating conditions rather than used in a general form.
Physical AI differs from consumer-facing generative AI because it must account for how environments behave and what may happen next. In practice, that means teams need models that can learn physical relationships, generate data for rare or hard-to-reproduce scenarios, and support testing before systems are deployed in the real world.
World models help address one of the field's biggest constraints: the cost and difficulty of collecting enough real-world data. That challenge is especially acute for unusual events and edge cases that are hard to capture safely and repeatedly.
NVIDIA is making its Cosmos world foundation models available under the Linux Foundation's OpenMDW 1.1 licence, giving developers access to model weights and allowing them to adapt the models using their own data and computing infrastructure.
Model range
At the centre of the announcement is Cosmos 3, an open model family for physical AI. It is designed to combine scene understanding, synthetic data generation and action prediction in one family of models, rather than requiring separate models for each task.
The range includes Cosmos 3 Super with 64 billion parameters, Cosmos 3 Nano with 16 billion, and Cosmos 3 Edge with 4 billion. NVIDIA says the Edge version is designed for deployment on devices and can run across RTX graphics processors, DGX systems and Jetson hardware, including Jetson Thor.
According to NVIDIA, Cosmos 3 can be used as a vision language model, a world simulator that predicts future states, or a basis for world action models. The aim is to let developers use the same model family across data generation, simulation and post-training workflows.
NVIDIA also highlighted benchmark performance across a range of public tests. It says the model family ranks first on Artificial Analysis for open-weights text-to-image and image-to-video generation, on PAI-Bench for world generation, and in the image-to-video category of Physics-IQ. It also says it ranks first on RoboLab for robot policy, while Cosmos 3 Super is the highest-ranked open model on VANTAGE-Bench for vision understanding.
Simulation tools
The software side of the strategy relies on Omniverse libraries and OpenUSD. NVIDIA says these tools help developers build simulation-ready environments and manage 3D assets and sensor configurations across digital twins, simulation work and synthetic data generation.
This approach is meant to reduce repeated work when teams alter environmental conditions, sensors or assets during development. For physical AI programmes, such changes can be frequent as systems are tuned for different operating settings.
NVIDIA positions openness as a practical requirement rather than simply a licensing preference. A general model, it argues, will not have seen a developer's exact robot, vehicle, sensors or work setting, so post-training and adaptation are necessary for deployment.
Industry uptake
NVIDIA cited a list of groups using Cosmos in different sectors. In robotics, it named Doosan Robotics, LG Electronics, Samsung Electronics and Skild AI. In autonomous vehicles, it pointed to Li Auto, Xiaomi and Afari.
For vision AI, NVIDIA says Centific, Fogsphere, Linker Vision, Milestone Systems and Yuan are using Cosmos 3 for industrial AI and smart spaces applications. These uses range from AI agents for video analysis to systems intended for operational environments.
Beyond Cosmos, NVIDIA says its broader physical AI software stack includes Isaac GR00T for robotics, Alpamayo for autonomous vehicles and Metropolis for vision AI. The company is using the open-model push to tie those efforts together around a common approach to world modelling and simulation.
NVIDIA also pointed to the Cosmos Coalition, which brings together model builders, developers and other groups working on physical AI. The company says the coalition contributes models, research and evaluation methods, and has recently expanded to Japan to support work on factories, logistics, agriculture, construction, healthcare and transportation.
The move reflects a wider debate in AI over open versus closed models, but in NVIDIA's case the argument is closely tied to industrial deployment. In physical AI, the issue is less about broad access alone than whether companies can tailor a model to a narrow operational task using their own data.
By combining open model weights, simulation software and benchmark claims, NVIDIA is making the case that world models will form a shared base layer for robots, vehicles and vision systems. It says open world models are already being used to generate training data, test policies and specialise physical AI systems.