Runway is researching technology to generate video in real time, aiming to replace current step-by-step prompting with a live, interactive streaming experience.

Runway is working on a new approach to AI video generation that allows users to stream video as they describe it, rather than waiting for finished clips. The company aims to minimize the time required to generate the first frame and enable users to actively steer the video output. This research seeks to address the inefficiency of current models, which require users to wait for results and restart the process if the output is incorrect.

Technical Approach and Model Development

The research builds upon Runway's GWM-1, a "General World Model" introduced in December 2025 that utilizes the Gen-4.5 architecture. Unlike standard models that build frames sequentially, which can cause errors to compound, Runway is training its model on its own outputs. This method teaches the system to correct its own deviations rather than amplifying them. The company also demonstrated a research preview at Nvidia's GTC conference in March, where a model running on the Vera Rubin platform delivered the first frame in under 100 milliseconds.

Economic and Practical Applications

Runway argues that real-time generation is more cost-efficient because faster models consume less GPU time, potentially making previously unprofitable applications viable. The company identifies education, gaming, and robotics as primary long-term use cases. For robotics, Runway has introduced GWM Robotics to generate synthetic training data. Other companies are pursuing similar strategies; for instance, Waymo uses a world model based on Google Deepmind's Genie 3 to simulate rare road scenarios for its autonomous driving fleet.