Runway Releases Solaris Interface World Model, Real-Time Interactive Interface Generated Frame by Frame After User Actions
Runway has announced its first "Interface World Model" Solaris: the system no longer calls pre-written web or application code, but generates the next state of an interactive interface frame by frame in real-time after user actions such as clicking and dragging; Runway positions this as an operating system capability that generates applications and websites during use.
Solaris treats the visual display itself as the application: a single world model simultaneously handles rendering and user interaction without going through DOM, CSS, JavaScript, or other intermediary code representations. Every frame seen by the user is both an interface and an interaction state; this means the interface does not need to be fixedly constructed by engineers first and then changed through subsequent version updates.
The system is based on the transformation of Runway's Gen-4.5 video generation model and continues the path of the GWM-1 general world model released in December 2025. The model has been adjusted to understand interactive inputs, generate images sequentially in an autoregressive manner, and reduce the generation steps required for real-time inference through step distillation and self-training.
Runway states that the development goals for Solaris include real-time interaction, consistency across complete sessions, and 720p visual quality. In its disclosed user research, 250 participants completed nearly 7,500 pairs of comparisons; Solaris outperformed the code-generated Claude Opus 5 interface with a 61% success rate in instruction compliance comparisons and a 71% success rate in natural interaction comparisons. These results were conducted by Runway itself and have not been independently verified.
Runway also used GPT-4o, Gemini 2.5 Pro, and Fable 5 as control objects in reconstruction tasks, employing SSIM and DINOv3 visual similarity metrics, claiming that Solaris outperforms leading large language models in structural similarity and information retention. The company has not yet disclosed complete benchmark details, model weights, source code, or API.
In terms of market mechanisms, Runway aims to shift the software value chain from "developers writing functions - front-end displaying functions" to "models synthesizing interfaces on demand." If this path holds, inference computing power, world models, and high-quality interaction data will become the main beneficiaries; traditional front-end development, templated SaaS interfaces, and low-complexity design delivery will face price pressure. However, at this stage, buyers are still constrained by accuracy, auditability, and accessibility compliance, and critical enterprise processes cannot solely rely on visually generated interfaces.
Solaris is still in the early research and partner testing phase, and Runway has only opened early access applications. Disclosed limitations include insufficient text stability, potential inaccuracies in interface content, long-session consistency issues, and incomplete support for APIs relied upon by accessibility technologies such as screen readers.
Source: Public Information
ABAB AI Insight
Runway's previous product path did not start from traditional software development tools: the company first established models, data, and inference infrastructure with generative video, then expanded capabilities into the GWM-1 general world model in December 2025. Gen-4.5 addresses continuity between video frames, GWM-1 extends this continuous prediction to environmental states; Solaris transforms "next frame prediction" into "next interface state prediction after user actions." This is a technological migration route from content generation, scene simulation to interactive software generation.
The key to capital and resources is not to replace a few front-end engineers, but to converge the design, interaction logic, and rendering required for software production into high-cost model inference. Solaris's public information shows that Runway is collaborating with NVIDIA to build computing infrastructure suitable for real-time frame-by-frame inference; this means that value is more likely to flow first to GPU suppliers, cloud computing power, model training data, and platforms with distribution entry, rather than simply to application packaging vendors at the prompt level.
Historically, Figma moved interface design collaboration from local design files to the cloud, while Webflow and Bubble transformed part of front-end delivery into visual building; they still rely on deterministic components, databases, and rule engines as their underlying layers. The difference with Solaris is that it eliminates the conversion layer from visual interface to code implementation, instead outputting interface states via generative models. It is currently closer to "generative prototypes and personalized front ends," rather than replacing systems that require deterministic transaction execution like financial transactions, medical records, or enterprise ERP.
This is essentially an attempt at pricing power transfer: traditional software charges based on functional modules, seats, and versions, while if generative interfaces can be stably implemented, the pricing unit will shift to each intent understanding, state generation, and real-time inference. The premise for this change is that the model can continuously map user inputs to correct, verifiable, and accessible states; therefore, factual accuracy, long-term consistency, and accessible interfaces are not marginal engineering issues, but decisive constraints on whether the model can enter commercial software budgets from the demonstration layer.
ABAB News · Cognitive Laws
Once the interface is generated, code becomes a base cost of the product
Automation first lowers execution prices, then competes for decision entry
Computing power determines the upper limit, credibility determines cash flow