NVIDIA ships Cosmos 3 Edge, a 4B world model for on-device robots
NVIDIA open-sourced Cosmos 3 Edge, a 4B-parameter world model that reasons and generates robot actions locally, hitting 15 Hz real-time control on Jetson Thor.
NVIDIA released Cosmos 3 Edge on July 20, a 4-billion-parameter open world model designed to let robots and vision AI agents perceive, reason and act entirely on-device — no datacenter round-trip required. Weights are available on Hugging Face under the Linux Foundation's OpenMDW-1.1 license.
Two towers, one model
Cosmos 3 Edge pairs an autoregressive reasoning tower (with a 2B dense reasoning component) and a diffusion generation tower in a Mixture-of-Transformers design, linked through shared multimodal attention. The architecture works in both directions: given an action, it predicts the visual consequences; given observed changes, it infers the actions that caused them. Actions map into a unified representation spanning camera motion, autonomous vehicles, robot arms and humanoids, so one model covers embodiments that previously needed separate policies.
The "edge" designation is literal. The model runs across NVIDIA's device lineup — RTX PRO and GeForce RTX GPUs, DGX systems and Jetson modules, including the new Jetson T2000 and T3000. On Jetson Thor it generates 32 actions per inference and sustains real-time control at 15 Hz at 640×360 resolution, with video generation supported at 256p and 480p from 12 to 30 fps. NVIDIA claims the top spot on VANTAGE-Bench in the 4B class, and says developers can fine-tune locally on a GeForce RTX 3070 or better. A post-trained manipulation policy, Cosmos 3 Edge Policy (DROID), ships alongside for pick-and-place tasks.
Why it matters
Robotics has been waiting for its "small model" moment: world models were previously too heavy to run where robots actually operate. A permissively licensed 4B model with real-time control rates puts world-model-driven autonomy within reach of hobbyist GPUs and embedded modules, not just lab clusters. It also tightens NVIDIA's grip on the physical-AI stack — the model is open, but the deployment targets are Jetson — and raises the bar for Xiaomi, Figure and the growing field of embodied-AI labs betting that data and deployment scale, not parameter count, will decide the robotics race.