Runway releases Praxis-1, a world action model with open weights
A 27B-parameter mixture-of-experts diffusion transformer under Apache 2.0; Runway reports 46.2% on NVIDIA's RoboLab-120 against 42.9% for FLUX 3 Action and 28.0% for pi0.5.
- Models & capabilities
- Open weights & ecosystem
- Minor
Runway, the video-generation company, released the weights of Praxis-1, a “world action model” it introduced in September as its first. Praxis-1 is a 27-billion-parameter mixture-of-experts diffusion transformer that predicts video and robot actions together, built on the same large-scale video pretraining as Runway’s general world models. Runway released the weights and the inference and fine-tuning code, covering full fine-tuning and LoRA, under the Apache 2.0 licence, which it said includes commercial use.
Runway’s central claim is a benchmark result. On NVIDIA’s RoboLab-120 leaderboard it reported 46.2 per cent success across 1,200 episodes, against 42.9 per cent for FLUX 3 Action, 36.8 per cent for Cosmos3 Nano and 28.0 per cent for pi0.5, and said that makes it the top-performing world action model on the board and ahead of every open-weight vision-language-action model. The table is Runway’s, dated November 2026, and the gaps between the top entries are a few points on one simulation benchmark. In a second test, Runway said a Praxis-1 checkpoint fine-tuned on Google’s RT-1 robot, which was absent from its training data, reached 80.0 per cent on SimplerEnv against 52.4 per cent for the original RT-1 policy.
The post argues that learning from general video gives the model a head start on new robot bodies, citing Runway’s earlier research that web video matches teleoperated robot video as pretraining data. Launch partners Noble Machines, Standard Bots and Ultra, Runway said, had been running it on a legged humanoid, a six-axis industrial arm and a dual-arm warehouse robot. The weights are on Hugging Face with LeRobot support and on GitHub, and the model runs in vLLM-Omni, with hosted inference on Reactor. Runway said it runs on NVIDIA hardware from data-centre GPUs to the Jetson Thor module on the robot.
Runway framed the choice as a policy position, saying that US leadership in physical AI depends on openness and interoperability that American models do not yet offer for physical use cases, and that it had committed to open weights when it announced the model. It described the release as building on its work as a founding member of the Cosmos Coalition, and said it and NVIDIA are opening technology stacks for physical-AI developers. The company did not report real-world success rates in this post, and the sources reviewed do not include independent evaluations of the model.