Many MindedRunway opens the weights on a 27B model that teaches robots to act
Runway AI Inc. has released Praxis-1, a world action model for robotics, with open weights under the Apache 2.0 license, which permits commercial use. The company first announced it in September with launch partners Noble Machines, Standard Bots and Ultra. They ran it on a legged humanoid, a six-axis industrial arm and a dual-arm warehouse robot.
Praxis-1 is a 27B-parameter mixture-of-experts diffusion transformer that predicts video and actions together, built on Runway's video foundation model. On NVIDIA's RoboLab-120 benchmark it scored 46.2% success over 1,200 episodes, against 42.9% for FLUX 3 Action, 36.8% for Cosmos3 Nano and 28.0% for π0.5. It ranks fifth overall on the leaderboard. In a Noble Machines plank-picking task it learned from half an hour of examples and succeeded 15 of 15 times. On a messy unseen pile it cleared the whole pile 6 of 10 times. Fine-tuned on Google's RT-1 robot, which was absent from its training data, it reached 80.0% on SimplerEnv versus 52.4% for the original RT-1 policy.
The release includes weights plus inference and fine-tuning code, with full fine-tuning and LoRA supported. It is on GitHub, Hugging Face (with native LeRobot support) and vLLM-Omni, and hosted inference is available on Reactor. It runs on NVIDIA hardware from H100 and B200 to RTX 5090, RTX PRO 6000 and Jetson Thor. Runway also says its research found web video matches teleoperated robot video as pre-training data for policy performance.
For abundance, the point is the barrier. A developer can now start from a capable, commercially usable robot model instead of training one from scratch, and learning a task from half an hour of examples points toward cheaper robot adaptation for physical work.
Caveats: the benchmark results and the claims of being the top-ranked world action model and open-source end-to-end policy come from Runway, not independent verification. The 6 of 10 result shows messy, unseen scenes remain hard. Runway says one checkpoint adapted to multiple robot bodies with light fine-tuning from partners, but gives no details. Its stated next focus is scaling real-world deployments with partners.
Source: The Robot Report