Meta launches Llama-3 Small and Llama-3 Vision with design-focused prompts

AI · 5 min read

Meta today announced Llama-3 Small, a compact language model optimized for low-latency creative workflows, and Llama-3 Vision, a multimodal model trained to parse UI screenshots, wireframes, and design specs.

Both models ship with curated prompt libraries and example notebooks specifically for tasks like generating component copy, producing accessibility hints, extracting layout hierarchies from screenshots, and creating design token suggestions. Meta also published a set of design-oriented evaluation benchmarks to measure layout understanding and UI semantics.

The company emphasized on-device deployment and parameter-efficient fine-tuning for teams that need private, fast-inference models. Early partners — including tooling vendors and a handful of design studios — reported improved generation quality for microcopy and component naming compared with general-purpose LLMs.