MENU

Runway Praxis-1: 27B Parameter Robot AI Model and Specs

XenoSpectrum USA
Overview
Runway has open-sourced Praxis-1, a 27 billion parameter robot control AI that leverages video-prediction knowledge for action generation under the Apache 2.0 license. While simulation success rates reached up to 80.0% after fine-tuning, the model requires significant computational resources and robot-specific data for real-world deployment.
In Depth

Key Findings

Runway has released the model weights and inference/training code for Praxis-1, a “world action model” designed to generate robot movements based on instructions and camera feeds. The model is provided under the Apache 2.0 license, allowing for commercial use and custom fine-tuning. Performance evaluations in simulation showed success rates of 80.0% for an RT-1 model in SimplerEnv and 46.2% for a DROID model in RoboLab-120, though these results were achieved only after robot-specific additional training.

Technical Details

Praxis-1 consists of a video-prediction foundation model paired with an action-generation module of approximately 314 million parameters, totaling 27 billion parameters. The system processes instructions, camera images, and robot posture to generate actions—typically 24 steps (1.6 seconds) at 15Hz, with 21 steps for the initial inference.

  • Hardware Requirements: Standard bf16 weights are 54.4GB, requiring a single 80GB GPU. Inference on an H100 takes approximately 0.55 to 0.6 seconds.
  • Optimization: A quantized NVFP4 version reduces weight size to 17.5GB, maintaining a success rate of 46.5% compared to 47.8% for the bf16 version in specific tests.
  • Training Pipeline: The process involves general video pre-training, followed by training on videos of human hands and robots, joint training of video prediction and action generation, and final robot-specific fine-tuning.

Background and Context

Prior to the general release on November 11, 2026, the model was trialed by Noble Machines, Standard Bots, and Ultra. To achieve the reported simulation results, Runway utilized 128 NVIDIA H100 GPUs for 50,000 updates, using approximately 277 hours of successful demonstration data for the DROID model.

Significance and Outlook

Praxis-1 allows developers to utilize general video knowledge rather than building models from scratch. However, Runway notes that the model still struggles with ambiguous instructions, similar-looking objects, and complex multi-step tasks without individual prompts for each stage. Future research will focus on long-term action planning and self-improvement capabilities.

Source: https://xenospectrum.com/runway-praxis-1-open-robot-model/

Get our weekly technology intelligence — free

Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.

Subscribe Free — Weekly Tech Intelligence

By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.

  • Your email and selected fields are used only to deliver the newsletter.
  • We never share your information with third parties.
  • You can unsubscribe anytime via the link in each email.

See our Privacy Policy for details.

Takes about a minute · Unsubscribe anytime

Published by Troy-Technical, an independent site run by one engineer with a career in materials development.
About the author / Contact info@troy-technical.jp
Let's share this post !

Author of this article

TOC