Runway gives away Praxis-1, its robot brain, as open weights

Runway said on Nov. 11, 2026, that Praxis-1, the world action model it introduced in September to control robots, is now available to everyone with open weights under the Apache 2.0 license, including for commercial use, on GitHub and Hugging Face.

Robot data is the scarce thing in robotics, and Runway's whole pitch is that ordinary video can stand in for it. Giving the model away with a commercial license is a land grab: if robot makers fine-tune on Praxis-1, Runway becomes the default brain without building a single robot. The leaderboard lead is narrow and self-reported, so the real test is what developers get out of it on their own hardware.

Praxis-1 is what Runway calls a world action model. It predicts a video of what happens next and the robot's next moves together. Runway says it is built on the same large-scale video pretraining as the company's general world models. Pretraining is the long first lesson on a huge pile of video, before anyone adapts the model to one robot. Runway introduced Praxis-1 in September. On Nov. 11, 2026, it published the weights. Weights are the numbers the model learned. Those lines are Runway's.

The model is a 27-billion-parameter mixture-of-experts diffusion transformer. A parameter is one stored number. Twenty-seven billion is the size. Mixture-of-experts means only some of those numbers wake up for a given step, so a large model does not do all of its math every time. A diffusion model starts from noise and cleans it into an answer. A transformer is the network shape behind most modern AI. That description is Runway's.

Runway says Praxis-1 is the top world action model on Nvidia's RoboLab-120 leaderboard: 46.2 percent success across 1,200 episodes. An episode is one test run. On the same board, Runway puts FLUX 3 Action at 42.9 percent, Cosmos3 Nano at 36.8 percent, and π0.5 at 28.0 percent, with the data as of November 2026. Runway also says Praxis-1 beats every open-weight VLA. A VLA, a vision-language-action model, looks, reads a command, and outputs a robot move. Open-weight means anyone can download the numbers and run the model. Those scores are Runway's own figures. An outside lab has not signed them.

Runway says it fine-tuned Praxis-1 on Google's RT-1 robot, a robot that was not in the training set. Fine-tuning is extra training for one machine after the general lesson. On SimplerEnv, a simulated test of that robot, Runway reports 80.0 percent success, against 52.4 percent for the original RT-1 policy. A policy is the program that picks the robot's next move. Those two percentages are Runway's.

Launch partners Noble Machines, Standard Bots, and Ultra ran the model on three kinds of machine: a legged humanoid, a six-axis industrial arm, and a dual-arm warehouse robot. A six-axis arm is a factory arm with six joints. Those partners and those three robots are the ones Runway named.

The weights, and the code to run the model and to fine-tune it, are under the Apache 2.0 license, which allows commercial use. Apache 2.0 is a license that lets a company use, change, and sell what it builds on the model. Runway includes full fine-tuning and LoRA. LoRA is a smaller add-on training that does not rewrite every number. The files are on GitHub and on Hugging Face. Hugging Face has native support for LeRobot, a common toolkit for robot learning. Runway also points to vLLM-Omni, a program that serves the model to many users at once, and to hosted inference on Reactor, which means running it on Runway's computers instead of yours. The announcement is at runway.com/news/company-news/praxis-1-is-now-available. Those lines are Runway's.

Runway says the model runs on Nvidia hardware from the large H100 and B200 servers down to a desktop RTX 5090 or RTX PRO 6000, and on a Jetson Thor, a small Nvidia computer that can sit on the robot itself. That hardware list is the company's.

Runway's research claim is that ordinary web video matches teleoperated robot video as pretraining, when you measure how well the resulting policy performs. Teleoperated means a person drove the robot by remote control while a camera recorded it. Runway also says it is a founding member of the Cosmos Coalition with Nvidia. Both claims are the company's.

The picture is two white-and-black robot arms lifting a dark red cloth toward a black paper shopping bag on a white table. Runway says the arms are running Praxis-1. The credit is Runway, a still from the company's demo.

In plain terms, the company known for generative video gave away the robot model it introduced in September. Anyone can take the weights from GitHub or Hugging Face and use them commercially under Apache 2.0. The leaderboard lead is Runway's own reading of Nvidia's board, and it is narrow: 46.2 percent against 42.9 percent for the next model.