Runway unveils Solaris to render software UIs without code

Runway has introduced Solaris, a new interface world model designed to generate software interfaces on the fly rather than rendering conventional code. Instead of executing HTML and CSS to display web pages, Solaris processes user actions like clicks, drags, and keystrokes as conditioning inputs to predict the next visual frame directly. By taking recent visual history as context, the system transforms browser tabs into real-time simulations. Behind the interactive visual front end, traditional large language models and autonomous agents handle long-lived logic, state maintenance, and tool calls.

To achieve smooth interaction, Runway envisions a hybrid architecture where a lightweight local model renders frames at high frame rates while a heavier cloud model periodically provides deeper context. Today, the approach faces key technical bottlenecks, specifically persistent memory across interactions and latency, where even half a second of delay ruins the feel of an application interface. Beyond software interfaces, Runway is exploring playable environments through GWM Worlds 2, moving toward a future where specialized variants for software, games, and robotics eventually converge into a single foundation model.

The underlying technology is already delivering measurable results in physical domains. Runway reported testing robot policies within its world model simulations and achieving a 0.95 correlation with real-world policy performance rankings. Based on these developments, Runway's Anastasis places a 3-year bet on world models entering production robotics while simultaneously generating a meaningful share of everyday software interfaces dynamically tailored to individual user preferences for layout, style, and density.

For software developers and AI practitioners, Solaris represents a paradigm shift from building static user interfaces to training interactive generative systems. Rather than manually coding responsive visual components, engineering teams may soon focus on designing conditioning pipelines and balancing local inference latency against cloud model context. If latency constraints are overcome, front-end development could transition from writing explicit UI markup to guiding visual foundation models that adapt instantly to user inputs.