Runway releases Praxis-1 robot model under Apache 2.0

Runway released Praxis-1, its first robot model, with open weights under the Apache 2.0 license on November 11.

The announcement is in a company post and a technical write-up, and the weights and code are live on Hugging Face and GitHub. Runway first announced Praxis-1 in September as its first world action model.

What changed is the release itself. Runway shipped three checkpoints, one fine-tuned on the DROID dataset, one on the bimanual YAM robot, and the base mid-trained model, along with the model code, a policy server, a MuJoCo simulation and code for both full and LoRA fine-tuning. Apache 2.0 covers commercial use.

Praxis-1 is a 27-billion-parameter mixture-of-experts diffusion transformer that predicts video and robot actions together. It runs on the same video pretraining as Runway's general world models, reads instructions through a frozen Qwen3.5-2B text encoder and works in the latent space of a frozen Wan2.2 VAE. Give it camera observations, the robot's state and a task, and it predicts the actions to carry the task out. Runway says its research shows web video matches teleoperated robot video as pretraining data.

On Nvidia's RoboLab-120 benchmark, Runway reports 46.2% success across 1,200 episodes, ahead of FLUX 3 Action at 42.9%, Cosmos3 Nano at 36.8% and π0.5 at 28.0%. The Robot Report's Brianna Wessling noted it ranks fifth overall on that leaderboard, behind two language-model agents (86.0% and 57.1%), Lakeside (51.0%) and BiMind v0.5 (49.2%). Runway's narrower claim, that Praxis-1 is the top world action model and the top open-source end-to-end policy, holds on the same page. Fine-tuned on Google's RT-1 robot, which never appeared in its training data, Praxis-1 reaches 80.0% on SimplerEnv against 52.4% for the original RT-1 policy.

"We view open world models as a compounding advantage," Andy Chen, Runway's head of robotics, told The Robot Report. Runway's launch partners are Noble Machines, Standard Bots and Ultra. With Noble Machines, Runway says Praxis-1 learned a warehouse task from half an hour of examples and placed wooden planks 15 times out of 15.

A lab best known for generating video is now shipping control for real machines, and it is betting that video pretraining is the cheaper route to a general-purpose robot.

Praxis-1's weights and code are free to download on Hugging Face and GitHub, with hosted inference on Reactor.