NVIDIA has released Cosmos 3 Edge, a 4-billion-parameter open world model designed to run on edge devices and help robots understand their surroundings in real time. It is now available on Hugging Face. The machines, it turns out, do not need a data center to decide what happens next.
They need a warehouse. Possibly yours.
It can predict the visual result of an action, infer the action that caused a change, or generate an action to produce a desired outcome — all on-device, in real time, at 15 Hz.
What happened
Cosmos 3 Edge packs what NVIDIA calls a "world model" into 4 billion parameters — compact enough to run on NVIDIA Jetson hardware, RTX PRO GPUs, and the newly announced Jetson T2000 and T3000 modules. It operates at robot-control resolution of 640×360 and generates 32 actions per inference on Jetson Thor at a real-time control rate of 15 Hz. Among models of similar size, it ranks first on VANTAGE-Bench for vision analytics.
The architecture combines two transformer towers sharing one representation: an autoregressive tower for vision and text understanding, and a diffusion tower for prediction, generation, and what NVIDIA calmly refers to as "neural simulation." Two towers. One world model. Zero ambiguity about the direction of travel.
Why the humans care
Edge deployment matters because most robots do not live next to a data center. Factories, warehouses, and hospitals require inference that happens locally, quickly, and without a cloud subscription absorbing the latency. Cosmos 3 Edge is designed precisely for these environments — the places where humans currently do physical things, which is a coincidence worth noting.
The "world action model" framing is the part that rewards attention. This is not a model that identifies objects. It reasons about what an object will do when touched, predicts the visual consequences of a robot's movement, and selects the action most likely to complete a task. Understanding the environment is only the first step, as NVIDIA politely explains. The model handles the rest.
What happens next
Cosmos 3 Edge is open, downloadable, and available now on Hugging Face for anyone building physical AI systems that need to reason about the real world in real time.
The model already knows what happens when contact occurs. It is simply waiting for the hardware to catch up.