Runway Releases Open Weights for Praxis-1, Its Video-Trained Robot Control Model

The company introduced Praxis-1 in late September but at first gave access only to select partners. Its release post, dated November 11, says the open licence covers the model weights as well as the inference and fine-tuning code, and that the weights are on Hugging Face and the code on GitHub. Runway also published a research report with new benchmark results.

The release matters because Praxis-1 is an attempt to carry what video models learn from ordinary footage into robotics, where training data has traditionally been expensive to collect. Making it open lets robot makers run and modify it on their own hardware rather than through an API. "Open weights give hardware developers the flexibility to run, inspect and adapt a world action model on their own terms," Runway said.

Runway describes Praxis-1 as a 27-billion-parameter mixture-of-experts diffusion transformer that predicts video and robot actions together, a design it calls a world action model. The model takes in camera images, text instructions and the robot's position, predicts how the scene will change, and works out the movements needed. It keeps adjusting as new camera frames arrive.

The model started as a video model trained on everyday footage and was then fine-tuned on videos of human actions and on robot data. It still needs some robot-specific data. To adapt it to a Franka robot arm, Runway used about 277 hours of recordings from the public DROID research dataset, according to its report, and between 1,400 and 2,600 hours for a dual-arm robot. Whether the approach needs less robot data overall than models without general video pretraining is still unclear.

On Nvidia's RoboLab-120 leaderboard, which tests sorting, grasping and stacking across 120 simulated tasks, Runway reports a 46.2% success rate for Praxis-1. That compares with 42.9% for FLUX 3 Action, 36.8% for Cosmos3 Nano and 28.0% for π0.5. The result puts Praxis-1 first among world action models but fifth on the overall leaderboard. On a real Franka arm, Praxis-1 succeeded in 88 of 100 attempts across ten tasks, against 80 for FLUX 3 Action. After fine-tuning on about 95 hours of data from the mobile robot Google used to build RT-1, which was not in Praxis-1's earlier training, it reached 80% success in simulation, compared with 52.4% for Google's original RT-1 policy. All of these tests were run by Runway and have not been independently verified.

Since September, robotics companies Noble Machines, Standard Bots and Ultra have been testing the model on a legged humanoid, a six-axis industrial arm and a dual-arm warehouse robot. Runway says it runs on Nvidia hardware ranging from H100 and B200 data centre GPUs to RTX 5090 workstation cards and the Jetson Thor module mounted on robots. It supports full fine-tuning and LoRA, and it is also offered through vLLM-Omni and hosted inference on Reactor. The GitHub documentation says the standard version needs a GPU with 80 GB of memory. A compressed version should fit within the RTX 5090's 32 GB, according to Runway's estimates, though peak memory use has not yet been measured.

Runway's report also lists the model's limits. Praxis-1 sometimes confuses objects that look alike, struggles with vague instructions and cannot reliably plan long sequences of actions by itself. Runway positions it as a fast executor of movements that could in future work under general-purpose AI models that break complex jobs into individual steps.