Runway, the company best known for its AI video tools for filmmakers, has released Praxis-1, a robot-controlling AI model, with its weights open to everyone under the Apache 2.0 licence — a move that lets any robotics developer or researcher download, run and adapt it, including for commercial use.
Praxis-1 is what Runway calls a “world action model”: a 27-billion-parameter mixture-of-experts diffusion transformer that predicts the next few frames of video and, at the same time, the physical actions a robot should take to bring that future about. It is built on the same large-scale video pretraining behind Runway’s general world models, and it was first introduced in September before being tested with selected partners.
The headline number comes from Nvidia’s RoboLab-120 leaderboard, which covers 120 tabletop tasks with ten trials each. Runway reports that Praxis-1, fine-tuned on the DROID robot dataset, completed 555 of 1,200 episodes — a 46.2% success rate — placing it first among world action models and open-source robot foundation models, ahead of FLUX 3 Action at 42.9%, as well as π0.5 and Nvidia’s GR00T.
The release includes everything needed to run and adapt the model on a developer’s own hardware: the weights, inference code and full or low-cost LoRA fine-tuning code, available on GitHub and Hugging Face with native support for the LeRobot robotics library, plus integration with the vLLM-Omni serving engine. Runway also offers hosted inference through its Reactor service.
Three checkpoints are included: a base mid-trained model, a version fine-tuned on DROID, and one fine-tuned for the bimanual YAM robot. The complete model contains 27.2 billion parameters and occupies 54.4 GB in bfloat16 format, allowing it to run on a single 80 GB GPU. Runway says it runs out of the box across Nvidia hardware, from data-centre GPUs such as the H100 and B200 to workstation cards such as the RTX 5090 and RTX PRO 6000, and on the Jetson Thor computer inside the robot itself.
In practice, Praxis-1 generates each 1.6-second chunk of actions in about 0.55 to 0.6 seconds on one H100, fast enough for real-time control. Runway says the model can adapt quickly to new robot bodies: fine-tuned on Google’s RT-1 mobile manipulator, which it had never seen during training, it averaged 80.0% success across four SimplerEnv tasks, against 52.4% for the original RT-1 itself.
Three launch partners have already run Praxis-1 on their own machines: Noble Machines on a legged humanoid, Standard Bots on a six-axis industrial arm, and Ultra on a dual-arm warehouse robot. Runway says one partner fine-tuned the model on just 154 teleoperated episodes — about 29 minutes of demonstrations — to get it working on a new robot.
The release matters because robotics has long been held back by a data problem: unlike language or images, robot behaviour is expensive to record, and the best models have mostly stayed locked inside large companies. By turning general internet video into a teacher for motion and object interaction, then aligning that knowledge with robot-specific training, Runway is arguing that video pretraining can shortcut the data bottleneck.
There are caveats. All of the performance numbers are company-reported and have not yet been independently verified, and open model weights do not automatically produce safe robot motion — developers still have to bind the model to their own safety limits and test it on their own hardware before it can be trusted around people.
Even so, the release is a significant moment for open robotics. Until now, the strongest robot foundation models have been proprietary; Praxis-1 is the first open-weight world action model to top a major public leaderboard, and it arrives at a moment when Google, Nvidia and others are all racing to build the general-purpose “brain” for humanoid and industrial robots.
For researchers and small robotics teams, the practical effect is immediate: a state-of-the-art robot policy that previously would have required a large lab’s budget can now be downloaded and fine-tuned on a single GPU. Whether that translates into robots that work reliably in homes and warehouses will depend on what the community builds with it in the months ahead.
Runway says the model is available from today on GitHub, Hugging Face and its own developer platform, and that it will continue to update the release as partners report results from real-world deployments.
