Runway, a company known mainly for video generation, has unveiled the Solaris model. It is the first model from a new group of systems it calls Interface World Models. Solaris does not create code that is then rendered by a browser. It renders the interface itself, frame by frame, in real time and based on what the user is currently doing. When you click a button, no event handler is triggered. The model simply draws the next frame showing what happened after the click.
Runway takes a different approach
Runway bases the entire concept on one problem. Every operating system, from old terminals to Linux and macOS, determines what appears on the screen and what happens when someone clicks it. Applications then run within the constraints imposed by the operating system and do not change until developers release an update. Solaris bypasses this model and renders the screen itself.
Software designs are more visual today than ever before. Image models can generate an entire screen that is almost indistinguishable from a finished product. But an image does not work like a website or application. Between the design and the finished product, there is still a translation into an intermediate language, usually code, which Runway says limits the possibilities on both sides. Every behavior must be described and programmed in advance, so the finished software is only a reduced subset of everything the user might want to do. At the same time, the visual richness of the original design is lost.
Solaris handles rendering and responses to input simultaneously. A single model generates every frame and every response to a user action. There is no conversion step, so according to Runway, nothing is lost and the entire frame becomes the interface.
Three things it does differently
The first is its purely visual nature. When the image is also the application, there is no need for a second, hidden implementation beneath what the user sees. Runway demonstrates this with a virtual clothing store where the showroom itself is the interface. All it needs is a single photo of the user as a reference. The user can then take a shirt from the rack, drag it onto themselves and try it on, or rearrange the entire display much like in a physical store.
The second feature is that the scene is alive. The application is rendered continuously, so it keeps changing instead of waiting for the next action. Reflections shift with the light, and objects react when someone moves them. The user can enter an instruction such as move the table so I can see what it looks like, or change the color of the sofa. Runway claims that the result feels less like clicking through prebuilt pages and more like moving through a living environment.
Another aspect is its open-endedness. Traditional interfaces can only do what developers anticipated when programming them. Solaris can support entirely different behaviors within the same scene because it interprets input at runtime. Runway shows the same initial image, an X-ray of a hand, responding to the same dragging action in two completely different ways.
The company is also rethinking the mouse itself. When interactions are described in natural language rather than through programming, they do not need to be determined in advance, and every object in the scene can become a tool. Click on a cat, and subsequent clicks transfer the color and texture of its fur to whatever you touch. Click on a painting, and you can start drawing in its style.
What Solaris is built on
Solaris is based on the Gen-4.5 video generation model and continues in the direction Runway established with the GWM-1 general world model. The company modified the model so that it understands interactions and responds immediately. Solaris treats user input as a condition for the next frame, much like text or an image. It tracks clicks, drags and other actions during generation and uses them as guidance for what should happen next. Because it only ever sees interactions that have already occurred, it learns the relationship between an action and its visual consequence. This means no one has to program what should happen after a click or drag.
According to Runway, speed was the biggest obstacle. Interaction stops feeling natural at a delay of around half a second, while conventional diffusion models need seconds or minutes to produce a clip. Runway solved this in three steps. First, it taught the model to generate frames sequentially, with each one depending only on the preceding context. It then shortened the lengthy denoising process to just a few steps. Finally, it trained a fast model on its own outputs so that the image would remain stable even during extended operation. According to the company, the same modification that made the model faster also made it orders of magnitude cheaper to run than a conventional diffusion model for video.
Decision-making and rendering are separate. Solaris draws the scene frame by frame, while a language model determines how the interface evolves. It reads the user's requests and decides whether an interaction should merely modify the current scene or transition to a new one. It also describes behaviors that keep the scene alive and composes the prompts Solaris uses to draw. The user provides an initial state, such as a branded environment or a scene featuring a product, and the model then streams frames in real time. The system contains no prebuilt screens or templates to return to.
Uses and weaknesses
According to the company, today's language models cannot handle even basic computer tasks, such as booking a hotel or buying groceries. This is because they are trained on programmed interfaces, so they learn a specific layout and cannot cope with a slightly different page, such as the website of another hotel. Solaris can provide an agent with an environment that is constantly changing, including layouts that have never existed before.
For now, the model struggles most with text. Stable, legible lettering is among the hardest tasks in video generation, while user interfaces rely on text more than almost anything else. One possible route is a combination in which image models render text-heavy screens when a short pause is acceptable, while the video model handles fluid interaction.
Another problem is reliability. In educational or commercial situations, a convincing but incorrect answer is worse than no answer at all. For now, Solaris stays grounded in what the user provides. The initial frame can be assembled from real product photos and other materials, anchoring the scene in objects that actually exist. The company is continuing to explore ways to connect generation to verified source materials during extended operation, including reference images, product data and documents. Long sessions remain problematic because it is difficult to keep both the image and its meaning consistent, as does accessibility, since the generated interface must also work with screen readers and similar assistive tools.
Solaris runs at 720p, and Runway is positioning it as the first step toward a new interaction layer. The company expects it to make the application no longer the basic unit with which the user works. After all, if the operating system can generate the interface itself, there is little reason to organize software into a fixed catalog of applications. According to Runway, a store no longer needs a fixed layout that everyone sees in the same way, and a tutorial does not need to present everyone with the same sequence of steps.
The model is not yet publicly available.
