4 billion parameters, and it runs on a single GPU.
That is the claim behind Cosmos 3 Edge, NVIDIA’s open-weight world model for what the company calls Physical AI. A "world model" means a single neural system that can represent and simulate environments, reason about outcomes, and generate actions across modalities. "Open-weight" means the model weights are available for researchers and developers to download and run locally. "Edge" means it can operate on devices outside the cloud, like robots or vehicles.
NVIDIA’s product page presents Cosmos 3 as a multimodal omnimodel that handles text, image, video, sound and action. Independent reporting from Gigazine and NationPress captures the new Edge variant: roughly 4 billion parameters, open-sourced, and tuned to run on a single graphics card. NVIDIA also published distilled image and video generators, Cosmos3-Super-Text2Image and Cosmos3-Super-Image2Video, the same day.
This changes the calculus for physical systems. Until now, many powerful world models demanded large, remote clusters. A model that fits a single GPU lowers the hardware barrier for on-device autonomy. Teams that prototype robots, autonomous shuttles, or smart infrastructure can run the model locally for lower latency and for data privacy. Open weights let researchers audit, adapt, and distill the model further for constrained hardware.
The release also signals a maturity in tooling. Distilled image and video models paired with a compact world model create a path to richer on-device perception, simulation, and action planning. That matters for developers, because closed-loop testing no longer needs constant cloud roundtrips. It also matters for operators, who can reduce cloud costs and avoid connectivity risk.
Those gains come with trade-offs. Running sophisticated world models in the real world raises safety and verification needs. Autonomous vehicles and industrial robots require rigorous benchmarking and standards before wide deployment. The evidence today is strong that the model exists and runs locally. We still need public benchmarks, robustness tests, and deployment reports to judge real-world safety and reliability.
Why this matters: Cosmos 3 Edge makes a powerful, multimodal world model practically available to on-device robotics and vehicle teams. That accelerates experimentation and could shorten the path from lab prototypes to deployed autonomy.
If the model performs as described, the next year will test whether local world models deliver predictable, auditable behavior outside the lab. Expect rapid iteration, and a new set of questions about how we test and certify autonomous systems that no longer depend on remote servers.