NVIDIA Open-Sources Cosmos 3 Edge AI Model to Power Smarter Robots and Edge Devices

NVIDIA has unveiled Cosmos 3 Edge, a new open-source world model designed to help robots and vision AI systems better understand, predict, and interact with the physical world in real time.

The model is now available on Hugging Face's Cosmos 3 repository, enabling developers to build AI applications for factories, warehouses, hospitals, and other edge environments.

The 4-billion-parameter model is optimised for memory-constrained edge devices while delivering data center-level performance. It supports NVIDIA's edge computing platforms, including RTX PRO GPUs, GeForce RTX GPUs, DGX systems, and Jetson modules, including the newly announced Jetson T2000 and T3000.

Unlike traditional vision-language models, Cosmos 3 Edge functions as both a vision AI model and a World Action Model (WAM), enabling robots to understand their surroundings, reason about changing environments, and generate actions in real time.

Running at robot-control resolution (640×360), the model can generate 32 robot actions per inference and achieve 15 Hz real-time control on NVIDIA Jetson Thor.

According to NVIDIA, Cosmos 3 Edge ranks first on VANTAGE-Bench among comparable 4-billion-parameter models for vision analytics and delivers state-of-the-art performance for robot policy learning.

The model combines two transformer architectures—an autoregressive tower for reasoning over vision and language, and a diffusion tower for predicting video, audio, and actions. This shared representation enables AI systems to understand current scenes, simulate future outcomes, and determine how actions influence the environment.

NVIDIA is also releasing Cosmos 3 Edge Policy (DROID), a robot manipulation model trained for pick-and-place tasks, along with post-training scripts. Developers can fine-tune the models using NVIDIA H100 GPUs or DGX systems before deploying them to edge devices, expanding the use of physical AI across robotics, automation, and smart infrastructure.