NVIDIA drops a 4B world model that runs robots locally — no cloud required

The gap between a data center AI and a robot on a factory floor is usually measured in milliseconds of latency and gigabytes of VRAM. NVIDIA just closed that gap. Cosmos 3 Edge, a 4-billion-parameter world model released July 20, is designed to run entirely on-device, handling real-time understanding, reasoning, and action generation for robots and vision AI without a trip to the cloud.

This isn’t a scaled-down toy. It’s the third and smallest tier in the Cosmos 3 family, sitting alongside the 16B Nano and 64B Super models. Under the hood, it uses a Mixture-of-Transformers architecture with two towers that share multimodal attention layers. One autoregressive tower handles reasoning over vision and text; a diffusion tower handles prediction and generation over vision, audio, and action tokens. The trick is that both towers align their understanding through those shared layers. The model can look at a scene, reason about it, and then generate a video prediction of what happens next — and the action required to make it happen.

Here’s where it gets practical. Cosmos 3 Edge maps wildly different physical systems — a camera’s 9D motion vector, a single robot arm’s 10D action space, a humanoid’s 29D — into a single geometric representation. This lets the model connect pixel changes to control commands. It doesn’t just predict video; it predicts video as a result of an action. And it runs both ways: infer the effect of an action, or infer the action from its effect. NVIDIA also shipped a fine-tuned manipulation policy called Cosmos 3 Edge Policy (DROID), post-trained on pick-and-place tasks, with scripts developers can run on a modest H100 cluster or a DGX Station.

Deployment targets are aggressive. The model delivers 15 Hz real-time control on a Jetson Thor, generating 32 actions per inference from 640×360 observations. It also hits 12–30 fps generation at 256p and 480p resolutions. NVIDIA is pitching a GeForce RTX 3070 as a prototyping on-ramp, with post-training for a custom embodiment taking roughly a day. This is a direct shot at making physical AI something you iterate on locally, not something you rent from a hyperscaler.