A four-billion-parameter World Model runs on edge devices

The NVIDIA Cosmos 3 Edge model computes physics and motion on local boards. What action analysis changes for factory robotics.

Cosmos 3 Edge, released by NVIDIA on Hugging Face, is an open four-billion-parameter World Model designed for machines operating outside the data center, in factories, warehouses, or embedded in vehicles. It joins a family whose Nano and Super versions weigh sixteen and sixty-four billion parameters, respectively. Distribution is handled under the OpenMDW 1.1 license, including weights, code, and post-training recipes, for both commercial and non-commercial use.

A World Model learns how an environment transforms over time by representing objects, movements, spatial relationships, and the effects of an action. For a robotic arm, recognizing an object is not enough: it must locate it, track the movement of the gripper, anticipate what will happen upon contact, and choose the gesture that leads to the desired outcome. The model can predict the visual result of an action, infer the action that caused a change, or generate the action that will produce the intended effect.

The architecture, called Mixture-of-Transformers, relies on two complementary towers. The first, autoregressive, generates text token-by-token for comprehension and reasoning. The second, diffusion-based, synthesizes images, video, audio, and actions through iterative denoising. Both share their multimodal attention layers, aligning language, video, audio, and action into a common representation. The ways of describing an action—such as end-effector pose, camera movement, or grasping state—are reduced to compact geometric vectors covering about ten robotic configurations.

In policy mode, the model predicts an action and its expected visual consequence. NVIDIA cites thirty-two actions generated per inference on Jetson Thor and real-time control at 15 Hz, claims first place on VANTAGE-Bench among open models of comparable size, and releases a manipulation policy trained on the DROID dataset.

The model card also details the limitations. Lacking an explicit physics simulator, 3D geometry, object permanence, contact dynamics, and physical laws are only approximated, resulting in objects that disappear or deform and unrealistic collisions. Reasoning can fail regarding object states, causal relationships, or temporal order, and quality drops outside the training domain. NVIDIA writes that these outputs should not be taken as physically accurate simulations or as safety-certified decisions.