Into the Omniverse: How Open World Models Push the Frontier of Physical AI
15:00 · August 6, 2026 · NVIDIA
In July, NVIDIA joined more than 200 companies and organizations in signing “Open Weights and American AI Leadership,” an open letter arguing that AI leadership will be measured not by any single frontier model but by whether an open ecosystem reaches every sector.
Summary
NVIDIA has released Cosmos 3, a family of open world models built on a mixture-of-transformers architecture that combines vision reasoning, world generation and action prediction in a single set of weights. The models are intended to give physical AI systems the ability to learn how environments behave, forecast subsequent states and generate synthetic training data that respects physical constraints. Three sizes are offered: Cosmos 3 Super at 64 billion parameters for high-fidelity simulation, Cosmos 3 Nano at 16 billion for efficient post-training, and Cosmos 3 Edge at 4 billion for on-device deployment on Jetson and RTX platforms.
Physical AI deployments require specialization to particular robots, sensors and operating conditions. Cosmos 3 is released under the Linux Foundation’s OpenMDW 1.1 license, allowing teams to inspect, modify and continue training the weights on their own data and hardware. This openness addresses the practical difficulty of collecting rare or hazardous scenarios at scale, because the models can produce diverse, physics-grounded synthetic data that varies lighting, weather, object placement and trajectories.
Complementary Omniverse libraries and the OpenUSD framework supply simulation-ready environments in which the models can be exercised before real-world deployment. Developers can therefore use a single model family as a vision-language system, a forward simulator or the backbone for learned robot policies, rather than maintaining separate networks for each function. Early adopters already include robotics firms such as Doosan and Skild AI, automotive teams at Li Auto and Xiaomi, and vision-AI integrators working on industrial monitoring.
NVIDIA positions Cosmos 3 within a broader physical-AI stack that also contains Isaac GR00T for humanoid robotics, Alpamayo for autonomous driving and Metropolis for spatial AI. The company has formed the NVIDIA Cosmos Coalition to coordinate further open model development and evaluation methods across these domains.
Why it matters
The release of open-weight physical AI models by a major player like NVIDIA significantly lowers the barrier to entry for developing advanced robotics and autonomous systems. This is highly relevant for the Dutch market, which features strong logistics, agriculture, and high-tech manufacturing sectors that can leverage these transparent, open-source tools for innovation.
