Runway releases Praxis-1 open weights
Runway has made its 27B-parameter Praxis-1 world action model publicly available under Apache 2.0, giving robotics developers a commercially usable open model that tops NVIDIA's RoboLab-120 leaderboard on Runway's own figures.
Runway announced that Praxis-1 is now generally available with open weights under the Apache 2.0 license, commercial use included. The company first introduced the model in September, as we reported, and said then that it would ship as an open model. Until now, access was limited to launch partners Noble Machines, Standard Bots and Ultra. They have been running 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. It is built on the same video pretraining as Runway's general world models. Runway says it is the top world action model on NVIDIA's RoboLab-120 leaderboard, with 46.2% success across 1,200 episodes. The next three models score 42.9% (FLUX 3 Action), 36.8% (Cosmos3 Nano) and 28.0% (π0.5). The company also says that after fine-tuning on Google's RT-1 robot, which was not in its training data, Praxis-1 reaches 80.0% on SimplerEnv, against 52.4% for the original RT-1 policy. All of these figures come from Runway.
The release includes the weights plus inference and fine-tuning code, with support for both full fine-tuning and LoRA. The model is on GitHub, on Hugging Face with native LeRobot support, and in vLLM-Omni. Hosted inference runs on Reactor. Runway says the model works out of the box on NVIDIA hardware ranging from H100 and B200 data-center GPUs to the Jetson Thor robot computer. NVIDIA's NeMo Automodel provides tutorials for post-training.
Runway framed the release around U.S. leadership in physical AI and its role as a founding member of the Cosmos Coalition.