📅 November 11, 2026 • Runway
Runway has released the weights of Praxis-1, the robot model it introduced in late September with access limited to select partners. The launch post "Praxis-1 Is Now Available" on runway.com confirms the model is now available to everyone under the Apache 2.0 license with commercial use allowed: any team can download, fine-tune and deploy it, including inside commercial products. For robotics teams the open release removes the access barrier that has determined who could experiment with the model since its September introduction.
Praxis-1 is what Runway calls a world action model: a 27B-parameter video foundation model turned into a robot controller. It takes camera images, text instructions and robot position data, predicts how the scene will change and computes the movements the robot should make. The underlying bet, per Runway, is that web video matches teleoperated robot video as pretraining data for policy performance — ordinary internet video is enough to teach a model how objects, hands and tools behave, without waiting for expensive robot-collected datasets.
The release is unusually complete. It includes the weights, inference and fine-tuning code (full fine-tuning and LoRA), a policy server and a MuJoCo simulation. Three checkpoints are on Hugging Face: Praxis-1 fine-tuned on DROID, Praxis-1 fine-tuned on the bimanual YAM robot, and the base mid-trained model. Code is on GitHub with native LeRobot support, the model is available in vLLM-Omni, and hosted inference runs via Reactor. Runway says it runs out of the box on NVIDIA H100, B200, RTX 5090 and RTX PRO 6000 workstation GPUs — and on Jetson Thor mounted on the robot itself.
The launch comes with benchmark data. On RoboLab-120 — 120 tabletop tasks for a DROID-style arm, 10 trials per task — the checkpoint fine-tuned on DROID completed 555 of 1200 episodes, a 46.2 percent success rate. FLUX 3 Action scored 42.9 percent on the same leaderboard. The DROID-fine-tuned checkpoint is built for exactly this arm configuration. Runway says that makes Praxis-1 the top world action model and the top open-source robot foundation model on that leaderboard at publication.
The caveats are stated plainly. Four entries scored higher overall on the same leaderboard — two of them agents that directly use frontier models rather than trained policies — per Runway's report. Per The Decoder, the model also occasionally confuses similar-looking objects and cannot yet reliably plan complex action sequences on its own. Runway is a founding member of the Cosmos Coalition with NVIDIA, which advances open world models for physical AI. For robotics teams the practical value is a full stack — checkpoint, simulation, deployment targets — usable commercially from day one.
Sources: runway.com — the-decoder.com — huggingface.co
Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky
