Alibaba open-sources Qwen-UI-Agent — a GUI agent you can run locally
2B and 8B weights ship on Hugging Face under Apache-2.0 — it reads screens and clicks across phone, desktop and browser, no per-action API meter.
Copy markdown
An open GUI agent that drives your screen
Alibaba's Tongyi-MAI team released Qwen-UI-Agent, a foundation model that reads a screen and performs real clicks, typing and swipes across phones, desktops and browsers from one runtime — and stops to ask before sensitive actions like payments.
2B and 8B weights, Apache-2.0, on Hugging Face
The 2B and 8B checkpoints are live on Hugging Face under Apache-2.0, so commercial use is fine; the 32B and 235B variants aren't out yet. The 8B is small enough for a single consumer GPU — a computer-use agent you own, with no per-action API bill.
Reportedly rivals the frontier on screen control
Author-reported scores put it at 82.1% on MobileWorld, 79.5% on OSWorld-Verified and 73.6% on WebArena — reportedly competitive with Opus 4.8, Gemini 3.1 Pro and GPT-5.6 Sol on GUI tasks. Treat these as the team’s own eval until third parties reproduce them.
Build this weekend: a local phone-automation agent
Point the 8B at an Android emulator or a desktop sandbox and you have a screenshot-and-click loop that can book a ride, fill a form or scrape an app UI — running entirely on your machine, costing nothing per step.
Elsewhere: a Rust LSP using ~100x less RAM
rust-analyzer creator matklad shipped Rust Glancer, an early language server he says needs roughly 100x less memory than rust-analyzer — a lighter code-intelligence backend for large Rust codebases.