Runway has released Solaris, which the company calls its first "interface world model." Unlike tools that help people write front-end code, Solaris doesn't produce code — it produces the picture directly. Every frame of the interface a user sees is generated live by the model, and actions like clicks and drags are fed back into the model as generation conditions, changing the next frame accordingly.
Runway has given two performance figures: interaction latency held under half a second, and resolution maintained at 720p. Those two numbers define what it can and can't do right now.
Clicks Treated as Conditioning Signals
Solaris is built from a world model paired with a language model. The language model reads the user's request and decides whether this action should modify the current scene or switch to a new one; the world model renders that decision. The whole pipeline advances frame by frame, and the model learned on its own from training what happens when you click a given spot — nobody wrote a callback function for any button.
That marks a clear fork from the mainstream approach to AI-generated interfaces. Over the past two years, the standard playbook for large models handling UI has been to generate HTML or React code and hand it to the browser to render, with the interface's behavior still governed by that code. Solaris swallows the entire rendering layer — the image on screen is the application itself.
A 61% Preference Rate
Runway published a set of self-reported results: on interface-reconstruction tasks, Solaris beat frontier large models on both structural similarity and information retention. A separate user study with 250 participants compared instruction-following, and Solaris was the preferred choice 61% of the time.
All of these figures come from Runway itself — there's no independent third-party evaluation yet. And 61% isn't an overwhelming margin either; roughly, out of every 10 comparisons, about 6 favored Solaris and 4 favored the control. For a research prototype that just shipped, that's a reasonable showing, but it's still a long way from being "ready to replace existing software."
720p Is a Hard Ceiling
Runway lists four limitations of its own: unstable text rendering, coherence that degrades over long sessions, factual accuracy not yet solid enough for commercial use, and accessibility APIs that haven't been wired in yet.
The first and the fourth are the most serious. Unstable text rendering means information-dense interfaces like tables, forms, and code editors aren't viable for now — get a single digit wrong and the user has no way to tell. Without accessibility APIs, screen readers can't read anything at all, because there's no DOM structure to parse on screen — just pixels.
The 720p resolution ceiling draws another line. It's enough for mobile apps, casual games, and demo-style interactions, but desktop productivity software and data dashboards turn blurry the moment you shrink the font size to fit.
From Video Models to an Operating System
Runway has spent the past two years shifting its focus toward world models, with its valuation pushed to around $5.3 billion, and its business narrative has moved from "AI video tool" to "simulating the world." Solaris is the first product on that path aimed squarely at software itself — where video models simulate the physical world, an interface world model simulates the layer of human-computer interaction.
Solaris is currently in early access, with an application form posted on Runway's site; the company says it will roll the product out publicly alongside a few key partners first. For teams elsewhere, the barrier in this direction is interaction data: for the model to learn what happens when you click here, it needs a large volume of real interface-interaction traces, and that kind of data takes far longer to accumulate than video does.
Sources: Runway official research page, CocoLoop, AlphaSignal. Interaction latency, the resolution ceiling, and the 61% preference figure from the 250-person user study have been checked item by item.