On 20 July NVIDIA released Cosmos 3 Edge - an open model with about 4 billion parameters that understands its surroundings and decides the robot's next move on its own, right on the device, with no cloud. It runs in real time on their chip, Jetson. For robotics that means less latency and less dependence on an outside connection.
- NVIDIA released Cosmos 3 Edge on 20 July 2026 - an open 'world' model with ~4 billion parameters for robots.
- Runs on the edge device, no cloud; per NVIDIA's data - real time at 15 Hz on Jetson Thor.
- Also open: a 'pick and place' variant, a lightweight version, and the training scripts.
A robot that waits for an answer from the cloud is a robot that lags. That's exactly the pause NVIDIA is trying to remove.
What matters here isn't the parameter count. What matters is where it runs. Four billion is a small model, and small is deliberate - so it fits on a chip, right inside the robot. It sees, decides and moves without waiting on a network. For you, if you're building with robots, the math shifts: less latency, less dependence on connectivity, more work offline.
And the fact that it's open isn't a minor detail. Open weights plus open training scripts mean you can bring the model to your own task, not just use it as given. Either you rent someone else's capability, or you make the tool your own. Here it's the second.
On paper it sounds fast and open. The real test is on the shop floor: whether 15 Hz holds up when the floor is slick and the box isn't exactly where the model expects it. We don't know that yet. But the direction is right - the mind staying with the robot, not in a server across the ocean.