Overview
Praxis-1 is Runway's world action model, first introduced in September 2026 and released as open weights under Apache 2.0 on November 11, 2026, available to everyone including commercial use.
The technical shape is a 27B-parameter mixture-of-experts diffusion transformer that predicts video and actions together. That combination folds a world model and a policy into one object: the model has a prediction of what happens next, and actions are generated on top of that prediction rather than regressed straight out of a policy network.
On results it tops NVIDIA's RoboLab-120 leaderboard at 46.2% success across 1,200 episodes. On the same board FLUX 3 Action reaches 42.9%, Cosmos3 Nano 36.8% and pi-0.5 28.0%. Those numbers share a leaderboard and an episode count, so they bear direct comparison, and Runway says it outperforms every open-weight VLA.
The release is unusually complete. Apache 2.0 covers weights, inference code and fine-tuning code, explicitly including commercial use, with both full fine-tuning and LoRA. It ships through four routes: a GitHub repository, a Hugging Face collection with native LeRobot support, vLLM-Omni, and hosted inference on Reactor. Hardware runs from data centre H100 and B200 through workstation RTX 5090 and RTX PRO 6000 to Jetson Thor on the robot itself, out of the box according to Runway.
Cross-embodiment adaptation has one measurement worth keeping. Fine-tuned onto Google's RT-1 robot, a platform that never appeared in its training data, Praxis-1 reaches 80.0% on SimplerEnv against 52.4% for the original RT-1 policy. That suggests it learned how objects behave and how tasks unfold, rather than memorising one machine's joint angles.
The argument underneath matters more. Runway's research reports that web video matches teleoperated robot video as pretraining data for policy performance. If that holds, the bottleneck shifts from the cost of collecting data to how much general video a model can learn from, and general video is effectively unlimited. It also explains the emphasis on open weights: progress in physical AI depends on openness at both the hardware and software ends.
The launch partners show three embodiments: a legged humanoid from Noble Machines, a six-axis industrial arm from Standard Bots, and a dual-arm warehouse robot from Ultra, all running the model on their own hardware. Runway is also a founding member of the Cosmos Coalition, pushing open world models for physical AI with NVIDIA.
Be clear that this is a runnable starting point, not a tuned production recipe. 46.2% is first place on a leaderboard and nowhere near enough for a real process, which is why Runway points at adapting to your own hardware and domain data. Anyone building a product still has to build the fine-tuning data and the eval set.
Key Features
- Predicts video and actions together: A 27B-parameter mixture-of-experts diffusion transformer that outputs actions while generating the next stretch of video, folding world model and policy into one model.
- Apache 2.0 open weights: Weights, inference code and fine-tuning code released together, explicitly permitting commercial use, with both full fine-tuning and LoRA supported.
- Tops RoboLab-120: 46.2% success across 1,200 episodes, against 42.9% for FLUX 3 Action, 36.8% for Cosmos3 Nano and 28.0% for pi-0.5 on the same board, ahead of every open-weight VLA according to Runway.
- Data centre to onboard: Runs on H100 and B200 through RTX 5090 and RTX PRO 6000 down to Jetson Thor on the robot, out of the box.
- Four distribution routes: GitHub repository, Hugging Face collection with native LeRobot support, vLLM-Omni, and hosted inference on Reactor.
- Measured cross-embodiment transfer: Fine-tuned onto the RT-1 robot, absent from its training data, it reaches 80.0% on SimplerEnv against 52.4% for the original RT-1 policy.
- Web video as pretraining data: Runway's research reports web video matches teleoperated robot video for policy performance, shifting the bottleneck from data collection to how much general video a model can learn from.
- Validated on three embodiments: Noble Machines' legged humanoid, Standard Bots' six-axis industrial arm and Ultra's dual-arm warehouse robot all run it on their own hardware.
Use Cases
- Fast policy fine-tuning for your own arm
- Control for legged humanoids and dual-arm warehouse robots
- Replacing expensive teleoperation data with web video pretraining
- On-device inference on the robot itself
- Embodied AI research and reproduction
- Integration with the LeRobot ecosystem
- A policy baseline for industrial loading and sorting
Pros
- 27B that predicts video and actions together, folding world model and policy into one
- Apache 2.0 covering weights, inference and fine-tuning code with commercial use permitted and both full and LoRA tuning available
- 46.2% to top RoboLab-120, on a sample of 1,200 episodes that supports comparison
- Runs from H100 to RTX 5090 to Jetson Thor, including local inference on the robot
- 80.0% on SimplerEnv after fine-tuning onto an RT-1 robot outside its training data
- Four clear access routes through GitHub, Hugging Face, vLLM-Omni and Reactor
Pricing
Weights and code are free under Apache 2.0 with commercial use permitted, including full fine-tuning and LoRA. Hugging Face offers native LeRobot support, and Reactor provides hosted inference as a paid third-party service.
Summary
Very few open-weight robotics models run on the robot itself. Praxis-1 is a 27B mixture-of-experts diffusion transformer that predicts video and actions together, with Apache 2.0 covering weights, inference and fine-tuning code for commercial use, running from H100 to RTX 5090 to Jetson Thor onboard. It tops the leaderboard at 46.2% and reaches 80.0% on SimplerEnv after fine-tuning onto an RT-1 robot absent from its training data. The claim worth judging is the one underneath: web video can replace teleoperated robot video as pretraining data, and if that holds, the cost structure of robot data changes.
Version History
- Praxis-1 released as open weights (2026-11-11): Runway released Praxis-1 under Apache 2.0, covering weights, inference and fine-tuning code with commercial use permitted. The model is a 27B-parameter mixture-of-experts diffusion transformer predicting video and actions together, topping NVIDIA's RoboLab-120 at 46.2% against 42.9% for FLUX 3 Action, 36.8% for Cosmos3 Nano and 28.0% for pi-0.5. It runs on H100 and B200 through RTX 5090, RTX PRO 6000 and Jetson Thor. Distribution covers GitHub, Hugging Face with native LeRobot support, vLLM-Omni and hosted inference on Reactor. Fine-tuned onto the RT-1 robot, absent from its training data, it reaches 80.0% on SimplerEnv against 52.4% for the original RT-1 policy.