NVIDIA has introduced Cosmos 3 Edge, a compact world model with 4 billion parameters optimized to run directly on edge devices, such as the Jetson Thor platform.

What Happened
NVIDIA developed Cosmos 3 Edge using a Mixture-of-Transformers architecture. This architecture combines an autoregressive reasoning mechanism with a diffusion generator. When running on the Jetson Thor platform, the model provides an update rate of approximately 15 Hz at 640x360 resolution, generating action chunks to control robots.
Context
The development is aimed at transitioning from heavy cloud computing to local physical intelligence. The use of a hybrid architecture allows the model to effectively predict the physics of the surrounding world, which is essential for real-time operation.
Why It Matters for the Industry
For the industry, this signifies a shift from cloud-based control to local autonomy. The emergence of compact models capable of operating in a real-time control loop creates the foundation for commercially viable autonomous humanoids and industrial manipulators that function without dependency on cloud communication latency.
Why It Matters for Users
For end users, this is a step toward creating truly autonomous home and industrial robots. Such devices will be able to react instantaneously to changes in their environment, ensuring safer and faster interaction without the need for a constant internet connection.
Sources
Author
Look at AI, Editorial Team
![Nvidia Puts a 4B World Model on the Robot [Weekly Physical AI Roundup]](/assets/tg-news-media/f8/f83da5fa3330734d6fddb64f841db8ece331ffe608f989a0551e3bde1ba14dc9.jpg)