OpenAI unveils Mosaic-3: a multimodal model tuned for UI generation

AI · 4 min read

OpenAI today released Mosaic-3, a 220-billion parameter multimodal model that the company says has been specifically fine-tuned on UI screenshots, component libraries, and annotated design-to-code pairs. Mosaic-3 accepts images, design prompts and context tokens to output layout suggestions, accessibility annotations, and component code in both React and Flutter formats.

The model integrates a new “design intent” conditioning layer that helps the model prefer accessible color contrasts, consistent spacing, and semantic HTML patterns. OpenAI also shipped a companion CLI and a plugin for popular design tools so teams can iterate between mockups and production-ready components without leaving the design environment.

Early user tests reported that Mosaic-3 reduces initial UI prototyping time by 40–60%, with the biggest gains on mobile form flows and admin dashboards. OpenAI emphasized responsible defaults: color-contrast checks, keyboard navigation scaffolding and an explanations panel that surfaces why the model picked particular semantic tags.