The competition in AI for embodied intelligence has reached another milestone. Recently, well-known AI company Runway announced that its flagship world action model, Praxis-1, is now officially open source for developers and researchers worldwide under the permissive Apache 2.0 license. It has released all model weights, along with inference and fine-tuning code, and fully supports commercial use.
Top-Tier Performance: 27B-Parameter MoE Architecture Tops Authoritative Robotics Leaderboard
As a 27-billion-parameter Mixture-of-Experts Diffusion Transformer (MoE Diffusion Transformer), Praxis-1's core innovation is its ability to predict video and actions simultaneously. Previously, through collaborations with hardware partners such as Noble Machines, Standard Bots, and Ultra, Runway successfully ran the model on legged humanoid robots, six-axis industrial robotic arms, and dual-arm warehouse robots.
In NVIDIA's authoritative RoboLab-120 benchmark, Praxis-1 demonstrated dominant performance. Across 1,200 evaluation episodes, it achieved an overall success rate of 46.2%, surpassing strong competitors such as FLUX3Action (42.9%), Cosmos3Nano (36.8%), and π0.5 (28.0%), ranking first among world action models.
In addition, in transfer generalization tests, after lightweight fine-tuning on the Google RT-1 robot (which never appeared in its pre-training data), Praxis-1's success rate in the SimplerEnv simulation environment soared to 80.0%, far higher than the original RT-1 policy's 52.4%, fully demonstrating its strong first-mover advantage and cross-embodiment adaptability.
Hardware Compatibility and Ecosystem Support: Covering Everything from Data Centers to Edge Devices
At the practical deployment level, Praxis-1 shows extremely strong hardware compatibility. It integrates deeply with the NVIDIA ecosystem, supporting not only high-performance data center GPUs such as the H100 and B200, but also workstation-class graphics cards such as the RTX 5090 and RTX PRO 6000, and can even be deployed directly on edge-side chips such as the Jetson Thor carried onboard robots.
To help developers get started quickly, the model is now available on several mainstream code and hosting platforms:
Code and weights access: Developers can get the full content through GitHub and Hugging Face, and the Hugging Face version natively supports LeRobot.
Inference and fine-tuning: It supports invocation through vLLM-Omni and Reactor hosted inference; it also supports full fine-tuning and LoRA, so developers can easily adapt it to their own robot form factor, specific cameras, and specific tasks without retraining the entire model.
Advancing Open Physical AI: The Open Source Ecosystem Is the Inevitable Trend
Speaking about the motivation behind this full open-source release, Runway said that for the United States to maintain its lead in physical AI, it not only needs continued investment in foundational software and hardware, but also needs models to achieve a level of interoperability and openness that physical application scenarios currently lack.
