NVIDIA published Cosmos 3 Edge, a 4 billion parameter world model designed to run entirely on the compute board inside a robot, not in a data center rack. The weights went up on Hugging Face’s Cosmos 3 repository on July 20, alongside post-training scripts and a robot manipulation policy called Cosmos 3 Edge Policy (DROID). NVIDIA has already talked up Cosmos 3’s ability to link perception, prediction and action; this release is narrower and more consequential for builders, because it is the version small enough to fit inside a factory robot’s power and memory budget.
That distinction matters because most world models still assume a network connection. A robot that has to send camera frames to a cloud endpoint, wait for a prediction, and receive an action back is bound by round-trip latency and by whatever the network is doing that day. Cosmos 3 Edge is pitched at systems where that dependency is a liability: hospital robots, warehouse arms, and factory floor equipment where a dropped connection or a 200-millisecond lag can mean a missed grasp or a safety stop. Running the model locally also keeps camera and sensor data off the wire entirely, which matters for any deployment in a hospital or a facility with proprietary manufacturing processes.
Architecturally, Cosmos 3 pairs an autoregressive tower that handles vision and text for reasoning with a diffusion tower that handles vision, audio and action tokens for prediction and simulation. The two towers keep separate normalization layers but share multimodal attention, so the model can reason about a scene and then generate the video or action tokens that follow from that reasoning. Actions across different robot bodies (an arm’s end-effector pose, a gripper’s grasp state, a vehicle’s ego motion) get mapped into one geometric representation of translation, rotation and manipulation state, which is what lets a single checkpoint serve as either a reasoner or an action generator.
On Jetson Thor, Cosmos 3 Edge processes camera input at robot-control resolution and generates 32 actions per inference cycle while sustaining control at 15 Hz, fast enough to close a perception-to-motion loop without buffering. Among models of similar size, NVIDIA reports Cosmos 3 Edge ranks first on VANTAGE-Bench for vision analytics and leads on robot policy learning benchmarks. Those figures come from NVIDIA’s own release notes; the company has not published an independent third-party evaluation alongside them, so the ranking should be read as a vendor claim against a benchmark NVIDIA itself selected as the comparison point.
The edge release sits inside a broader hardware push. NVIDIA paired the model with two new Jetson modules, the T2000 and T3000, and Cosmos 3 Edge is explicitly tuned to run across that lineup as well as RTX PRO, DGX and GeForce RTX systems. That gives NVIDIA a matched stack: silicon built for edge robotics, and now a world model sized to actually run on it, rather than a data-center model developers have to compress themselves.
Open weights change who can use this. Because Cosmos 3 Edge ships on Hugging Face with post-training scripts, a robotics team can fine-tune it on a small cluster of H100 GPUs or a DGX Station and deploy the result to their own Jetson fleet without routing inference through NVIDIA’s cloud APIs or paying per-call. For teams already running perception on Jetson hardware, the near-term test is straightforward: benchmark Cosmos 3 Edge’s 15 Hz control loop against whatever policy model is currently running on-device, and check whether the accuracy gap NVIDIA claims on VANTAGE-Bench survives contact with a real production environment.
Reported by NVIDIA on July 20, 2026.
