A video company just handed every robot builder a free 27-billion-parameter brain

Most of this year’s robot-brain announcements have come with an asterisk. You can watch the demo, read the paper and admire the success rates, but you usually can’t download the thing. On November 11, Runway removed that asterisk for Praxis-1. The model’s weights, along with the code to run and fine-tune it, are now public under the Apache 2.0 license, which allows commercial use.

That is a notable move from a company most people know for generating video, not moving arms. Runway first showed Praxis-1 in September as its first world action model, and it built the model on the same large-scale video pretraining behind its general world models. Until this week, only a handful of launch partners had it.

What a world action model actually does

The robot models we cover most often, like Google DeepMind’s Gemini Robotics line, are vision-language-action models: they look at camera images, read an instruction and output movements. A world action model adds a step. It predicts what the scene will look like in the next moments together with the movements that would get it there, then acts on that prediction. Praxis-1 is a 27-billion-parameter mixture-of-experts diffusion transformer that predicts video and actions together.

Runway’s bet is that video is the cheap ingredient robotics has been missing. The company says its research shows ordinary web video works as well as teleoperated robot video for pretraining a robot policy. If that holds up outside Runway’s own tests, the bottleneck shifts away from paying people to puppet robots for thousands of hours and toward how much general video a model can learn from.

The numbers, and whose numbers they are

On NVIDIA’s RoboLab-120 leaderboard, Runway reports that Praxis-1 succeeded 46.2 percent of the time across 1,200 episodes. It lists FLUX 3 Action at 42.9 percent, NVIDIA’s Cosmos3 Nano at 36.8 percent and Physical Intelligence’s π0.5 at 28.0 percent. That makes it the top world action model on that board and ahead of every open-weight vision-language-action model, according to Runway.

Two caveats are worth keeping in view. A score under 50 percent on a hard benchmark is a ranking, not a guarantee the model can run a work cell on its own. And these are the company’s own reported results. The more practical figure may be the adaptation test: fine-tuned on Google’s RT-1 robot, which was never in its training data, Praxis-1 reached 80.0 percent on the SimplerEnv benchmark, against 52.4 percent for RT-1’s original policy.

Three very different robots already running it

Before the open release, launch partners ran Praxis-1 on their own hardware: a legged humanoid, a six-axis industrial arm and a dual-arm warehouse robot. The partners named are Noble Machines, Standard Bots and Ultra. That spread matters more than any single score. One checkpoint adapting to a walking robot, a factory arm and a two-armed picker, with light fine-tuning, is exactly the one-brain-for-many-bodies pitch that startups like Skild AI and Physical Intelligence have raised big rounds on.

Developers can adapt it with full fine-tuning or with LoRA, a lighter method that trains a small add-on instead of retraining the whole model. It is on GitHub and on Hugging Face with native support for LeRobot, the open robot-learning library, plus vLLM-Omni and hosted inference on Runway’s Reactor. It runs on NVIDIA hardware ranging from H100 and B200 data center chips to RTX 5090 workstation cards and the Jetson Thor computer that sits inside a robot.

Open versus closed, in one week

The timing makes the contrast sharp. DeepMind’s Gemini Robotics 2.5, which sorted parts on Mercedes-Benz lines with Apptronik’s Apollo 2, is going only to DeepMind’s robotics partners for now. NVIDIA unveiled GR00T 2 at the Conference on Robot Learning in Austin this week, and its robotics director Jim Fan said the weights will be open-sourced in December. Runway is shipping first.

Runway frames it as a matter of national competitiveness, arguing that U.S. leadership in physical AI needs American models that hardware makers can run, inspect and adapt on their own terms. Runway is also a founding member of the Cosmos Coalition for open world models and describes this release as part of joint work with NVIDIA, so it is less a break from NVIDIA than a release inside the same camp.

Our take

For a small robot company, a capable model you can download, inspect and fine-tune on a single workstation card changes what it costs to get started. The real test now moves out of Runway’s hands: whether independent teams can reproduce the leaderboard results, and whether anyone ships a product on Praxis-1 rather than another demo. Compare the bodies these brains run on in our humanoids guide, and if you want to work on robot learning, browse open roles on our jobs board.