KI
·
5 MIN
NVIDIA Cosmos 3: one open model for physical AI
NVIDIA Cosmos 3: one open world model for physical AI, folding vision, world generation, and action into one system.
LOCATION
Germany
SERIES
Design, Generative Media & Creative Pipelines
AUTHOR
Aashwin Shrivastava
PUBLISHED
NVIDIA used its Computex 2026 keynote to launch Cosmos 3, an open world foundation model for physical AI that folds vision reasoning, world generation, and action prediction into a single system. It is the clearest signal yet that the next frontier is not chat, but machines that understand and act in the real world.
01.
The shift: compute is data
The real world is effectively infinite, and real-world data is hard to scale. NVIDIA's framing is blunt: for physical AI, compute is data. Rather than collecting more footage, Cosmos 3 generates physics-accurate training data on demand, which is why NVIDIA says it cuts physical-AI training and evaluation cycles from months to days. It was trained on 20 trillion tokens of multimodal data, including close to a billion images and 400 million real and synthetic videos.
02.
One model, three jobs
Where earlier Cosmos releases split the work across separate models (Predict, Transfer, Reason, Policy), Cosmos 3 collapses them into one omnimodel built on a mixture-of-transformers architecture, with pixels, action, sound, and language flowing through a single autoregressive transformer:
As a VLM: it watches the physical world, describes scenes, and flags what matters.
As a world model: it generates physics-accurate synthetic video from an image, text, or video.
As a simulator: it closes the loop for robot policy training and evaluation.
For physical AI, compute is data. Cosmos turns the hardest part of any robotics project, generating enough realistic scenarios to train and validate against, into an API call.
03.
The coalition, and why we are watching
NVIDIA also launched the Cosmos Coalition, a collaboration with Agile Robots, Black Forest Labs, Generalist, LTX, Runway, and Skild AI to push next-generation world models forward together. The weights are open, and an 8B Nano variant is small enough to run locally, alongside a substantial technical report and developer guide.
This sits exactly where our AI, XR, and 3D-design work overlaps. A world model that generates physics-accurate synthetic environments is, in effect, a data and simulation engine, and the hard part of any robotics, digital-twin, or spatial-computing project is rarely the model, it is producing enough realistic scenarios to train and validate against. For clients building anything that has to understand or act in the real world, compute is data stops being a slogan and becomes a workflow.
Sources: NVIDIA Newsroom, NVIDIA Blog, NVIDIA Cosmos overview.
