MosaicAI ships Hermes: a 2B-parameter conversational design model with layout awareness

AI · 5 min read

MosaicAI says Hermes was trained on a curated corpus of design files, component metadata, and interface screenshots to learn spatial relationships and component semantics. Rather than only generating bitmap outputs, Hermes returns structured recommendations — component replacements, accessibility annotations, and layout constraints — suitable for programmatic application.

Early integrations include a Figma plugin and a REST API that outputs JSON patches compatible with popular UI frameworks. MosaicAI emphasizes low-latency inference through a hybrid quantization approach that keeps interactive feedback under 200ms for typical edits.

Design teams testing Hermes report faster iterations on accessibility fixes and content-aware layout shifts, while developers appreciate the model's ability to output actionable patch sets. MosaicAI plans a tiered release with on-prem and hosted options for enterprises concerned about IP and privacy.