Runway Open-Sources Robot Model Praxis-1: 27B Parameters Under Apache 2.0

What happened

Runway said on Nov 11 that robot model Praxis-1 is now available with open weights under the Apache 2.0 license, including for commercial use. Praxis-1 is a 27-billion-parameter mixture-of-experts diffusion transformer: it first predicts a short video of what happens next, then derives the robot's actions from that imagined future, an approach Runway calls a world-action model.

The release includes three checkpoints — the base model, a version fine-tuned on the DROID robot dataset and one fine-tuned for the bimanual YAM robot — plus inference code and full and LoRA fine-tuning code, on GitHub and Hugging Face, with native LeRobot support and vLLM-Omni, and hosted inference through Reactor. The model runs out of the box 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 itself.

Praxis-1 was first announced in September, when Runway committed to open-sourcing it; Nov 11 is the day the weights went public. The model first learns motion and object interaction from large-scale video, then is fine-tuned on robot data to map that knowledge onto physical actions. The full model is 54.4 GB in bfloat16 and needs a single 80 GB GPU; a 4-bit NVFP4 quantized version shrinks it to 17.5 GB.

Key facts

1. Architecture: A 27-billion-parameter mixture-of-experts diffusion transformer that predicts video and actions together
2. Benchmark: 46.2% on RoboLab-120 (555 of 1,200 episodes), ahead of FLUX 3 Action at 42.9%, Cosmos3 Nano at 36.8% and π0.5 at 28.0%; the top score among listed world-action models and open-source robot foundation models
3. Second benchmark: After fine-tuning on Google's RT-1 robot data, 80.0% average across four SimplerEnv tasks, against 52.4% for the original RT-1 model
4. Figures: All figures above are Runway's own, with no independent replication yet; 46.2% still means failing more than half of the simulated tabletop tasks
5. Release contents: Three checkpoints (base, DROID fine-tuned, YAM bimanual fine-tuned) plus inference and full/LoRA fine-tuning code; GitHub, Hugging Face with native LeRobot support, vLLM-Omni, and hosted inference on Reactor
6. Hardware and size: The full bfloat16 model is 54.4 GB and needs a single 80 GB GPU; the NVFP4 quantized version is 17.5 GB; runs on H100/B200, RTX 5090/PRO 6000 and Jetson Thor