Praxis-1: Runway's 27B MoE Diffusion Model Leads RoboLab-120 at 46.2% Success

Model Release and Benchmark Results

On November 11, Runway open-sourced Praxis-1 , a 27B-parameter mixture-of-experts diffusion model that simultaneously predicts the next video frame and the action to execute. The model weights, inference code, and fine-tuning code (including full and low-rank fine-tuning) are released under the Apache 2.0 license , permitting commercial use.

On the RoboLab-120 benchmark, Praxis-1 achieved a 46.2% success rate across 1,200 trials (data dated November 2026). This surpasses other listed models on the same table:

FLUX 3 Action : 42.9%

Cosmos3 Nano : 36.8%

π0.5 : 28.0%

Runway states that Praxis-1 is the highest-scoring world-action model on RoboLab-120 and exceeds all open-weight vision-language-action models.

Hardware Compatibility and Deployment

The release page notes that Praxis-1 runs on a wide range of hardware: from data-center GPUs ( H100 , B200 ) to prosumer cards ( RTX 5090 , RTX PRO 6000 ) and embedded robotics platforms ( Jetson Thor ). Inference is supported via GitHub, Hugging Face (with LeRobot integration), vLLM-Omni, and hosted inference on Reactor.

Demonstration Tasks

Demo videos on the release page play at real-time speed. Examples include:

A robotic arm placing a red garment into a black bag.

Placing a comb into a bag.

Inverting a cup onto a tray.

No per-task success rates are provided for these demonstrations.

Pre-training and Embodiment Transfer

Praxis-1 builds on Runway's general world-model line, using large-scale video pre-training. The release page explains that object dynamics and task progression are learned from web video. Without fine-tuning for specific hardware or datasets, the model already provides a starting point when transferred to a new robot body. The page also notes that web-video pre-training can match policy performance to teleoperation video, though no success rate is attached to that claim.

Fine-tuning Results on Unseen Robot

On a Google RT-1 robot not seen during training, Praxis-1 fine-tuned on SimplerEnv reaches 80.0% success , compared to the original RT-1 policy's 52.4% . This result is separate from the RoboLab-120 1,200-trial evaluation.

Real-world Deployments and Partnerships

Praxis-1 is already running on legged humanoids, six-axis industrial arms, and dual-arm warehouse robots. The page names Noble Machines , Standard Bots , and Ultra as partners. A single checkpoint, with light fine-tuning, has been adapted to several different robot morphologies.

Source

Runway announcement (November 11):

https://runway.com/news/company-news/praxis-1-is-now-available