Meta open-sources Llama-3-Design with layout tokens and style conditioning

AI · 6 min read

Llama-3-Design extends the Llama-3 family with a training corpus of annotated design artifacts, markup, and style guides. A key innovation is structured outputs: instead of free text, the model can emit JSON layout tokens, CSS variables, and component hierarchies that integrate more straightforwardly into design tools and code generators.

The repository includes conversion scripts to map model outputs to Figma, HTML, and mobile component formats. Meta also published evaluation benchmarks for layout fidelity and token alignment, encouraging community contributions to expand the dataset and improve safety filters for copyrighted UI assets.

Open-source maintainers cautioned that while the model lowers entry barriers, teams will still need domain-specific fine-tuning to match proprietary design systems. Meta pledged ongoing collaboration with the community and enterprise partners to develop best practices and governance patterns for deployed assistants.