NVIDIA opens Cosmos 3 Edge, a 4-billion-parameter world model that plans robot actions on a Jetson Thor

NVIDIA has pushed its world-model family down onto the robot itself. At SIGGRAPH on July 20, 2026 the company released Cosmos 3 Edge, a 4-billion-parameter open model that it says can reason about a scene and output robot actions in real time on a Jetson Thor computer, without a data-centre GPU.

Cosmos 3 Edge is the smallest of three Cosmos 3 models. NVIDIA launched the family at GTC Taipei on May 31 with the 64-billion-parameter Super and 16-billion-parameter Nano versions available immediately and Edge marked as coming soon. All three use what NVIDIA calls a mixture-of-transformers design that understands and generates text, images, video, ambient sound and actions. Edge pairs an autoregressive tower for reasoning with a diffusion tower for prediction, sharing attention layers across language, video, audio and action tokens. Its reasoning component is a 2-billion-parameter model based on Nemotron.

The weights are on Hugging Face, with inference and post-training code on GitHub. NVIDIA said Edge runs on Jetson modules, including the new Jetson T2000 and T3000, as well as RTX PRO, DGX and GeForce RTX GPUs. In BF16 precision the model occupies about 9 GB on a Jetson Thor. NVIDIA's technical blog reports that on a Jetson AGX Thor T5000 it generates a chunk of 32 future actions in about 1.53 seconds, and that each chunk covers roughly 2.13 seconds of robot motion at 15 Hz. The robot executes the start of a chunk and then replans, so a new plan is ready before the old one runs out.

For vision work, NVIDIA said Edge analyses 1080p video in as little as 22 milliseconds on RTX systems and about 30 milliseconds on L40 GPUs, and ranks first on the VANTAGE-Bench vision analytics benchmark in its parameter class. The company also released step-distilled versions of Cosmos 3 models that cut sampling from 50 denoising steps to 4, which it says makes inference 25 times faster with minimal quality loss.

The robot-control numbers are more modest, and NVIDIA publishes them. Post-trained on a dataset it calls Cosmos3-DROID, which contains 76,000 successful teleoperated trajectories, about 350 hours across 86 tasks and 564 scenes recorded on a Franka Panda arm with a Robotiq gripper, Edge reached 22.9% success across 120 language-conditioned manipulation tasks in closed-loop RoboLab evaluation. The baseline post-training run used 64 nodes of four GB200 chips each for 60,000 iterations, about 68 hours.

NVIDIA named Agile Robots, Doosan Robotics, Siemens and Skild AI as robotics companies evaluating Cosmos 3 Edge, and Centific, Vaidio and YUAN for smart-infrastructure vision agents. At the May launch it listed Agile Robots, Doosan Robotics, LG Electronics, Samsung Electronics and Skild AI as robotics adopters of Cosmos 3, and Li Auto in autonomous driving. Skild, which NVIDIA said in September uses Cosmos models and the Cosmos Curator tool in training its S1 model, is also a founding member of NVIDIA's Cosmos coalition alongside Agile Robots, Black Forest Labs, Generalist, LTX and Runway.

The target hardware matters to the strategy. NVIDIA's Jetson AGX Thor T5000, the module in its Isaac GR00T reference humanoid announced on June 1, delivers 2,070 FP4 teraflops with 128 GB of unified memory in a 40 to 130 watt envelope. A world model that fits in 9 GB leaves room on that module for perception, safety and a separate action policy. The reference humanoid itself is built on Unitree's H2 Plus and is expected from Unitree in late 2026.

The release sharpens a split in robot AI. Google DeepMind's Gemini Robotics 2, released ten days later, keeps its on-device action model with trusted testers. Skild AI offers S1 through an early-access sign-up. NVIDIA instead gives away weights for Cosmos 3 Edge and the 3-billion-parameter GR00T N1.7, which carries a commercial licence, and earns its return on the Jetson, RTX and data-centre hardware needed to train and run them.

In plain terms, a world model predicts what a scene will look like next. Cosmos 3 Edge uses that ability in two directions: it can generate or analyse video, and it can predict the motor commands that would move the scene towards a goal. Running both on the robot cuts the network delay of a cloud call and keeps camera footage on site, which matters for factories and for privacy rules in public spaces.

The limits are clear in NVIDIA's own figures. A 22.9% closed-loop success rate after 68 hours on 256 GB200 chips shows Edge is a research and prototyping base, not a deployable manipulation policy. A 1.53-second inference time is acceptable for slow tabletop tasks but not for fast dynamic motion. NVIDIA has not disclosed the licence terms for Cosmos 3 Edge in the materials reviewed, nor named any production deployment.