Runway releases 27b Praxis-1 world action model as open source under apache 2.0 license

Runway has made its flagship world action model, Praxis-1, publicly available for researchers and developers globally, releasing all model weights and associated code under the permissive Apache 2.0 license, allowing for commercial use. This 27-billion-parameter Mixture-of-Experts Diffusion Transformer excels at predicting actions and video concurrently. In NVIDIA's RoboLab-120 benchmark, Praxis-1 achieved a 46.2% success rate across 1,200 episodes, outperforming FLUX3Action (42.9%), Cosmos3Nano (36.8%), and π0.5 (28.0%). Furthermore, after minor tuning on the Google RT-1 robot, its success rate in the SimplerEnv simulation reached 80.0%, significantly exceeding the original RT-1 policy's 52.4%. Practically, the model supports deployment across diverse hardware, from H100 and B200 data center GPUs to edge chips like the Jetson Thor. Access is available via GitHub and Hugging Face, with support for vLLM-Omni and Reactor inference, alongside full fine-tuning capabilities. Runway stated that this open release is crucial for the United States to maintain its lead in physical AI by fostering necessary interoperability.