Runway releases Praxis-1 robot-control model with open weights and commercial-use rights

Robotics developers can now download and adapt the model Runway previously put in selected partners’ hands. On November 11, 2026, the company released Praxis-1 with open weights and code under Apache 2.0, including commercial use. The model predicts what a robot’s cameras will see next alongside the movements it should make.

From partner trials to public downloads

Runway introduced Praxis-1 in September. Its launch partners—Noble Machines, Standard Bots and Ultra—have been running it on a legged humanoid, a six-axis industrial arm and a dual-arm warehouse robot. The release announcement opens the same model to other developers and researchers, rather than introducing a separate successor.

The package covers the learned model weights, inference code for running it, and code for further training. It supports both full fine-tuning and LoRA adaptation. Developers have several routes to access it:

• Download through GitHub or Hugging Face, with native support in the LeRobot robotics software.
• Run it through vLLM-Omni, or use hosted inference on Reactor.
• Try the model through Runway Dev.

Runway lists support across NVIDIA hardware, from H100 and B200 data-center GPUs to RTX 5090 and RTX PRO 6000 workstation cards, plus Jetson Thor aboard a robot. Its technical specifications put the complete model at 54.4 GB in its standard numerical format, allowing it to run on one 80 GB GPU.

Imagining the next scene, then producing the motion

Praxis-1 is a roughly 27-billion-parameter model that predicts video and actions together. Runway calls it a world action model. Its technical report describes three connected parts: a video model that predicts the next scene, an action component that reads those predictions, and robot-specific components that translate different machines’ positions and movements into a shared format.

Each call takes an instruction, recent camera frames and the robot’s current pose. It returns predicted video and 1.6 seconds of control commands. The robot executes that chunk, sends back its camera observations and calls the model again. The action component reads the video model’s internal representations, not just a finished video clip.

Training starts with Runway’s general video foundation model. Further training adds footage of people and robots doing things, with first-person human video making up the largest share. Robot examples then supply action targets. Even after actions enter training, about 30% of tasks remain video-only, so the system continues learning from footage without matching robot commands.

The final training stage adapts the model to a particular robot. The shared video and action components carry forward their learned weights, while that robot’s input and output components start fresh. Runway reports 80.0% success on SimplerEnv after adaptation to Google’s RT-1, which was absent from earlier training, versus 52.4% for the original RT-1 controller.

46.2%Praxis-1

Runway reports 46.2% success across 1,200 evaluation episodes.

42.9%FLUX 3 Action

Runway’s comparison lists FLUX 3 Action at 42.9%.

36.8%Cosmos 3 Nano

Runway’s comparison lists Cosmos 3 Nano at 36.8%.

A category lead, not the overall lead

RoboLab-120 tests 120 simulated tabletop tasks for a DROID-style robot arm, using 10 trials per task. Runway says its DROID-fine-tuned Praxis-1 leads the leaderboard’s world action models and open-source robot foundation models. That is not an overall first-place result: its report identifies four higher-scoring entries, including two agents using frontier models directly and two proprietary systems.

Execution settings also involved a tradeoff. On a five-object pick-and-place test, Runway’s real-time chunking mode succeeded in 24 of 30 attempts, versus 28 of 30 for default chunked execution. Median completion times for successful attempts were 9.1 and 15.2 seconds, respectively. In that test, faster movement came with fewer successful attempts.