Ferret-UI Lite: Lessons from Building Small On-Device GUI Agents
Summary
Ferret-UI Lite is a compact 3B-parameter GUI agent designed to operate on-device across mobile, web, and desktop. The paper describes data curation from real and synthetic sources, chain-of-thought reasoning, visual tool-use, and RL with designed rewards to boost inference performance, achieving competitive grounding and navigation benchmarks on multiple datasets. The work highlights end-to-end GUI understanding on-device and discusses implications for edge AI and privacy-preserving automation.