Runway released the open weights for Praxis-1 on November 11, 2026. The release brings a robotics model previously used by launch partners to other teams that want to run it or adapt it to their equipment.
What Runway released
The release announcement describes Praxis-1 as a 27 billion parameter mixture-of-experts diffusion transformer that predicts robot actions and video together. Runway says the Apache 2.0 release covers the weights as well as code for inference and fine-tuning, including commercial use. The company lists GitHub, Hugging Face with native LeRobot support, vLLM-Omni and hosted inference through Reactor.
What the benchmark measured
Runway reports 46.2% success over 1,200 episodes on NVIDIA's RoboLab-120 benchmark. Its published table compares that result with 42.9% for FLUX 3 Action and 36.8% for Cosmos3 Nano on the same benchmark. A result from that benchmark describes its test conditions. It does not show how often the model will complete a task on a different robot, with different cameras or in a different workspace.
Runway also reports an 80.0% success rate on SimplerEnv after fine-tuning Praxis-1 for Google's RT-1 robot. That is a separate result with adaptation, not the score of an unmodified model on a new robot.
Where to start with a different robot
Runway names systems ranging from H100 and B200 data center GPUs to Jetson Thor on a robot. It says teams can adapt the checkpoint to their cameras, hardware and tasks using LoRA instead of updating every model parameter. Decide which test motions and failure cases matter on your own setup before choosing between hosted evaluation, local inference and an adapted checkpoint.
