Runway has open-sourced Praxis-1, its flagship world action model, releasing all model weights plus inference and fine-tuning code under the permissive Apache 2.0 license, according to AIBase.

The model is available now on GitHub and Hugging Face, where the Hugging Face version natively supports LeRobot, and developers can invoke it through vLLM-Omni or Reactor hosted inference.

Fine-tuning support covers both full fine-tuning and LoRA, so a team can adapt the model to its own robot form factor, cameras, and tasks without retraining the entire model.

Runway framed the move as a bid for openness in physical AI, arguing the US needs models to reach a level of interoperability that current physical application scenarios lack.

The surprise is both the release and the architecture. Praxis-1 is a 27-billion-parameter Mixture-of-Experts Diffusion Transformer whose core innovation, per AIBase, is predicting video and actions simultaneously in a single model.

Runway previously ran it with hardware partners Noble Machines, Standard Bots, and Ultra on legged humanoid robots, six-axis industrial arms, and dual-arm warehouse robots.

It also integrates deeply with the NVIDIA ecosystem, from H100 and B200 data center GPUs down to workstation cards like the RTX 5090 and RTX PRO 6000, and even edge-side Jetson Thor chips carried onboard robots.

A permissively licensed, commercially usable world action model at this scale upends the assumption that top robotics models stay closed.

AIBase reports that in NVIDIA's RoboLab-120 benchmark, Praxis-1 won first place among world action models with an overall success rate of 46.2% across 1,200 evaluation episodes, ahead of FLUX3Action at 42.9%, Cosmos3Nano at 36.8%, and π0.5 at 28.0%.

The transfer result may matter more: 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 reached 80.0%, far above the original RT-1 policy's 52.4%.

That cross-embodiment adaptability on an unseen robot is the kind of generalization result robotics builders have been waiting to see from an open-weight model.

With weights, code, and commercial rights all public, the immediate next step is developer experimentation: teams can pull Praxis-1 from GitHub or Hugging Face, fine-tune for their own robot bodies, and deploy from data-center GPUs down to Jetson Thor edge chips.

The release also lands amid a broader open-source push in AI tooling — Help Net Security reports that Amazon Web Services released AWS Emulator the same day, a tool simulating AWS services locally that AWS says was written entirely by autonomous AI agents.

If open world action models keep clearing benchmark bars, the moat in embodied AI shifts from holding weights to fine-tuning, data, and hardware integration.