Runway open-sources Praxis-1, a 27B world action model that controls real robots

Runway, best known for its generative video tools, has published the weights for Praxis-1, a 27B-parameter 'world action model' built to control real robots. The license is Apache 2.0 and explicitly covers commercial use, which makes this one of the largest permissively licensed robot-control models available. Runway introduced Praxis-1 in September with a commitment to open it, and today's release delivers on that promise.

A world action model extends the world-model idea that video labs have been pursuing. Instead of only predicting the next frames of a scene, it predicts how the scene responds to actions and outputs those actions. Runway's long investment in video prediction gives it a natural starting point: a model that has learned how objects move and interact in pixels can be adapted to choose motor commands that produce a desired future state. The coverage frames the release as a 'robot brain' rather than a narrow manipulation policy.

The timing matters. Physical AI is one of this week's busiest themes. Google DeepMind's Gemini Robotics 2.5 reported 96% task success in a Mercedes-Benz parts-handling trial, and NVIDIA keeps pitching its robotics and autonomy stack, including the open Alpamayo 2 Edge driving model. Most frontier robotics models remain closed or available only to partners. A 27B open model gives university labs and robotics startups a strong base to fine-tune on their own hardware without negotiating access, which is the same dynamic that drove Llama adoption in language models.

The caveats are real. The available coverage comes from a developer write-up rather than independent benchmarks, so there are no head-to-head numbers yet against Gemini Robotics or Physical Intelligence's models. A 27B model also needs serious onboard or tethered compute for real-time control, which limits where it can run. Watch for community fine-tunes on common arms and humanoids, published success rates on standard manipulation suites, and whether Runway pairs the weights with datasets or simulation tools. Those will show whether Praxis-1 becomes a shared foundation for open robotics or remains mostly a strategic statement from a video company widening its scope.