February 26, 2026
Tap wars on your home screen
Ferret-UI Lite: Lessons from Building Small On-Device GUI Agents
Tiny tap‑bot ignites “screenshots vs semantics” fight
TLDR: Ferret‑UI Lite is a tiny on‑device assistant that “looks” at screens and tries to tap through tasks, showing strong element detection but modest navigation wins. Commenters split between cheering Apple‑style small models and slamming screenshot‑based vision, arguing we should use accessibility labels instead for smarter control.
Ferret‑UI Lite is a bite‑size on‑device assistant that watches your screen like a human and tries to click its way through apps. The team says this 3B model (small by AI standards) learns by thinking step‑by‑step and trial‑and‑reward, and posts some eye‑catching scores—over 90% on spotting things on test screens and roughly 20–28% success on full navigation tasks. The paper is here if you like receipts: direct to paper.
But the real action? The comments. One camp is Team Tiny, cheering fast, local models that follow instructions inside your app. As one dev put it, they were “impressed at the speed and accuracy” and loved that it can translate a written instruction into tool use. That’s the Apple‑pilled crowd: keep it small, keep it on device, keep it useful.
Then the Accessibility Purists stormed in. “Why are we screenshotting the screen at all?” asks one commenter, blasting the “long way around” of visual recognition. Apple apps already have accessibility labels—basically a hidden map of the interface—and a throwback shout to Apple’s 1990s “Virtual User” had everyone dropping “back in my day” memes.
So the fight is set: see like a human vs. read the app’s map. Meanwhile, link‑droppers are tossing the paper like confetti, and everyone agrees on one thing: if tiny agents can actually navigate reliably on device, your phone might soon have a real tap‑happy co‑pilot.
Key Points
- •Ferret-UI Lite is a compact 3B on-device GUI agent operating across mobile, web, and desktop platforms.
- •The agent is trained with a curated mixture of real and synthetic GUI data, plus inference-time chain-of-thought and visual tool-use.
- •Reinforcement learning with designed rewards is used to improve performance and reliability.
- •Ferret-UI Lite achieves 91.6% (ScreenSpot-V2), 53.3% (ScreenSpot-Pro), and 61.2% (OSWorld-G) in GUI grounding.
- •Navigation success rates are 28.0% on AndroidWorld and 19.8% on OSWorld; related works Ferret-UI and Ferret-UI 2 expand UI understanding across platforms.