Runway has released the weights of Praxis-1, its robotics model, under the Apache 2.0 license, and the model now leads NVIDIA’s RoboLab-120 leaderboard among world action models. Anyone can download it, adapt it, and ship it in a commercial product. That puts a video-pretrained robot policy in the hands of hardware startups that cannot afford to train one.

Praxis-1 is a diffusion transformer with 27B parameters, built as a mixture of experts. It predicts video and robot actions together, and Runway says it learned how objects behave from large-scale video, the same pretraining family behind its general world models. The company introduced it in September. Launch partners Noble Machines, Standard Bots, and Ultra tested it on three bodies: a humanoid that walks on legs, an industrial arm with six axes, and a warehouse robot with two arms.

The leaderboard numbers are close. Runway reports that Praxis-1 succeeds in 46.2% of 1,200 RoboLab-120 episodes. FLUX 3 Action follows at 42.9%, Cosmos3 Nano at 36.8%, and π0.5 at 28.0%. The lead over the nearest rival is 3.3 points, and the gap to π0.5 is 18.2.

Read the top score soberly. Praxis-1 fails in more than half of the benchmark episodes, so this is a better starting point for robotics teams, not a finished worker. The benchmark belongs to NVIDIA, but the figures in the release come from Runway, and the announcement cites no independent reproduction.

The distribution matters as much as the score. Weights and inference code are out, along with training code for fine-tuning, which supports LoRA as well as updating every parameter. Developers can pull the model from GitHub, or from Hugging Face where LeRobot is supported natively. It also runs inside vLLM-Omni, and Reactor offers it as a hosted service. On the hardware side it covers NVIDIA data center GPUs such as the H100 and B200, workstation cards such as the RTX 5090 and RTX PRO 6000, and Jetson Thor, the computer that sits inside the robot.

The adaptation claim is the one builders should test first. Runway says it fine-tuned Praxis-1 on Google’s RT-1 robot, a machine absent from the model’s training set, and reached 80.0% success on SimplerEnv. The original RT-1 policy scores 52.4%. A single checkpoint has also adapted to several robot bodies with light fine-tuning, according to Runway. Its own research, the company adds, shows that footage scraped from the web performs as well as teleoperated recordings for pretraining a policy, which would shift the scarce resource from collected robot demonstrations to general video.

Runway frames the release as a policy position too. It argues that American leadership in physical AI depends on both hardware and software, and that American models are not yet open or interoperable enough for physical use. The company is a founding member of the Cosmos Coalition, which works on open world models, and says it is opening these stacks together with NVIDIA.

The timing sets up a clear comparison. Closed robotics models force a robot maker to rent intelligence by the call. An Apache-licensed model lets that maker own the fine-tuned result and run it on the robot itself.

Robotics teams choosing a base policy now have a commercially licensed option that leads the public board, and a LoRA run on their own robot will show within days whether the 3.3-point margin survives contact with their cameras and tasks.

Reported by Runway in its Praxis-1 release announcement, November 2026.