Robotics teams can now download Runway’s Praxis-1 and use it in products. The company has published the model’s weights and code under the permissive Apache 2.0 licence, two months after first showing it in September.
The model has 27 billion parameters and, in Runway’s terms, is a ‘world action model’ that forecasts video and the robot’s moves in one pass. Instead of learning from piles of robot recordings, it starts from the web-scale video Runway uses for its video generators and predicts the next frames and the robot’s actions together. Runway’s argument is that a model that has seen enough video already understands how things move and how a task plays out, so it needs less robot-specific data. Runway says its earlier research found that ordinary web video works as well as footage recorded by teleoperated robots for training a policy, so the bottleneck becomes the volume of everyday video available for pretraining.
On Nvidia’s RoboLab-120 leaderboard, Runway’s announcement puts Praxis-1 first among world action models at 46.2% success over 1,200 episodes. FLUX 3 Action follows at 42.9%, Cosmos3 Nano at 36.8% and π0.5 at 28.0%. Runway also says it outperforms every open-weight vision-language-action model. All of this is the company’s reading of the leaderboard as of November 2026.
Retargeting needs only light tuning, either of the whole model or through LoRA adapters, whether the change is a new robot body, new cameras or a new task. To show it, Runway adapted the model to Google’s RT-1 robot, which was never in the training data. On the SimplerEnv benchmark it scored 80.0%, where the original RT-1 policy managed 52.4%. Before the public release, other firms tested it on their own machines: a walking humanoid from Noble Machines, an industrial arm with six joints from Standard Bots and a two-armed warehouse robot from Ultra.
The hardware list is wide: H100 and B200 data-centre GPUs, RTX 5090 and RTX PRO 6000 workstation cards, and Jetson Thor, the computer that sits inside the robot. Code is on GitHub, the weights are on Hugging Face with LeRobot support, and Reactor offers hosted inference for anyone who prefers not to run it. Nothing depends on Runway’s own servers, so a lab can adapt and run it entirely on machines it controls.
Runway presents the release as part of an open, interoperable push for American physical-AI software, alongside Nvidia in the Cosmos Coalition. The announcement is dated 11 November, and it links tutorials for post-training the model with Nvidia’s NeMo Automodel.
