NVIDIA releases Cosmos 3 Edge, a 4B-parameter open world model for robots

Cosmos 3 Edge is NVIDIA's compact entry in physical AI: a 4-billion-parameter open world model published on Hugging Face's Cosmos 3 repository and built to run on the edge. It targets robots and vision agents that must perceive their environment, reason about it in real time and produce actions locally — without round-tripping to the cloud — with memory-efficient, high-throughput inference on NVIDIA edge hardware including RTX PRO GPUs and Jetson modules.

Releasing it as an open world model on Hugging Face is notable given the week's open-weights politics; NVIDIA is leaning into open distribution for robotics even as US policymakers debate restricting Chinese open models. It pairs with a Hugging Face + NVIDIA overview on 'The State of Simulation for Physical AI,' signaling a coordinated push on the embodied-AI tooling stack.

The strategic logic is to seed the robotics ecosystem on NVIDIA silicon at the edge, where latency and power constraints make small, efficient models essential. It also extends the Cosmos world-model line downward from data-center scale to on-device deployment.

Watch developer uptake on Jetson and whether Cosmos 3 Edge's reasoning holds up in real robotic deployments versus simulation benchmarks.