Runway released the weights of Praxis-1 on November 11, 2026, making its 27-billion-parameter world action model for robot control available to any developer under the Apache 2.0 license, including for commercial use. The company introduced Praxis-1 in September as its first world action model, and until now launch partners had been running it on their own robots. The release covers the weights as well as inference and fine-tuning code, with support for both full fine-tuning and LoRA.
Technically, Praxis-1 is a mixture-of-experts diffusion transformer that predicts video and robot actions together, built on the same large-scale video pretraining as Runway's general world models. The Hugging Face model cards describe a 27.2B-parameter policy made of a chunk-causal trunk and a 0.3B action expert, controlling a robot at 15 Hz with 24-step action chunks. Three checkpoints are published. The base model has no embodiment-specific heads and is meant as a starting point for fine-tuning rather than a ready policy. The DROID checkpoint targets a single Franka arm with an exterior and a wrist camera, and the YAM checkpoint targets a bimanual rig with two arms and three cameras. Both policy cards say inference needs one 80 GB GPU, and the GitHub repository mentions quantized NVFP4 versions and a checkpoint for the SO-101 arm that is still to come. Runway also lists support in LeRobot and vLLM-Omni, plus hosted inference on its Reactor service.
The headline claim is about benchmarks. Runway says Praxis-1 is the top-performing world action model on NVIDIA's RoboLab-120 leaderboard and outperforms every open-weight vision-language-action model, with 46.2 percent success across 1,200 episodes against 42.9 percent for FLUX 3 Action, 36.8 percent for Cosmos3 Nano and 28.0 percent for pi0.5. The Robot Report, which covered the release independently and spoke with Runway, added that the model ranks fifth overall on that leaderboard, so the comparison depends on how the field is cut. Runway also reports that after fine-tuning on Google's RT-1 robot, which was absent from its training data, Praxis-1 reached 80.0 percent success on SimplerEnv, versus 52.4 percent for the original RT-1 policy. The DROID card is more modest: 52.5 percent full prompt adherence on Runway's own evaluation set, with a 95 percent confidence interval of 35.8 to 68.3.
Runway's launch partners are Noble Machines, with a legged humanoid, Standard Bots, with a six-axis industrial arm, and Ultra, with a dual-arm warehouse robot. Andy Chen, Runway's head of robotics, told The Robot Report that with Noble Machines the model learned a plank-picking task from half an hour of examples and got it right 15 times out of 15, then cleared a messy pile of unseen planks in six of ten attempts. Those figures are Runway's account of partner trials.
The company frames the open release as a matter of physical-AI strategy. It argues that U.S. leadership will need open, interoperable models for hardware developers to run and adapt on their own terms, and it describes the work as part of the Cosmos Coalition effort with NVIDIA. Praxis-1 runs on NVIDIA hardware from H100 and B200 data center GPUs to RTX 5090 and RTX PRO 6000 workstation cards and Jetson Thor on the robot, and Runway points developers to NeMo Automodel tutorials for post-training.
Several limits apply. Benchmark and partner results are vendor-reported, and the leaderboard ranking is described differently by Runway and by The Robot Report. The frozen dependencies, the Wan2.2 VAE and the Qwen3.5-2B text encoder, are distributed separately under their own licenses. The base checkpoint cannot drive a robot without added heads. Runway says its next focus is scaling deployments with partners.