By Hamza Faisal
Runway, the company best known for AI video generation, has given away the brain behind its robots. The company has released Praxis-1, its 27-billion-parameter robot control model, with open weights under the Apache 2.0 license, meaning anyone can download it, run it, fine-tune it, and even use it commercially.
Runway first introduced Praxis-1 in September as its first world action model, built on the same large-scale video pretraining behind its general world models. Until now it had only been tested with selected partners. The public release on November 11 puts the full model, its inference code, and its fine-tuning code on GitHub and Hugging Face, with native support for the LeRobot robotics library and the vLLM-Omni serving engine. Hosted inference is also available through Runway's Reactor platform.
A video model that predicts what to do next
Praxis-1 works differently from most robot control systems. It is a mixture-of-experts diffusion transformer that predicts video and actions together: it imagines what the world will look like next, then turns that prediction into motor commands. The video backbone holds about 26.9 billion parameters, while a much smaller 16-layer action expert reads the backbone's internal state and produces the movements. Only around 2.8 billion backbone parameters are active per token, and the full model occupies 54.4 GB in bfloat16 format, so it fits on a single 80 GB GPU. With 4-bit NVFP4 quantization, the weights shrink to 17.5 GB.
Each call generates a 1.6-second action chunk in roughly 0.55 to 0.6 seconds on a single Nvidia H100, fast enough to keep up with a robot in motion. The model runs on Nvidia hardware from data center H100 and B200 chips to workstation RTX 5090 and RTX PRO 6000 cards, and even on the Jetson Thor computer mounted inside a robot itself.
Leading the benchmark — with caveats
On Nvidia's RoboLab-120 leaderboard, which covers 1,200 simulated episodes across 120 tabletop tasks, the DROID fine-tuned version of Praxis-1 completed 555 episodes, a 46.2% success rate. Runway says that is the highest result among listed world action models and open robot foundation models, ahead of FLUX 3 Action at 42.9%, as well as pi0.5 and Nvidia's own GR00T N1.6. The figures are Runway's own, and no independent tests have been published yet.
The model also adapts to hardware it has never seen. Fine-tuned on Google's RT-1 data, Praxis-1 averaged 80.0% across four SimplerEnv tasks, against 52.4% for the original RT-1 model. Three checkpoints ship with the release: the base mid-trained model, a version fine-tuned on the DROID dataset, and one fine-tuned for the bimanual YAM robot.
Already running on real machines
Three launch partners have already run Praxis-1 on their own hardware: 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 its robot working.
The release matters because capable robot foundation models have mostly stayed locked inside the companies that built them. By opening the weights, the training code, and the fine-tuning tools, Runway is betting that the fastest way to make robots useful is to let the whole robotics community build on the same brain — and to find out, in public, whether a model trained on internet video can really learn to handle the physical world.
By Hamza Faisal



