Qwen-UI-Agent tops GUI benchmarks. No weights yet.

Most of the software your business depends on has no API. The old POS, the supplier portal, the county permit site, the wholesale ordering page that still wants a fax number. A GUI agent — a model that reads a screen and clicks it — is the only automation path into that half of your stack. Alibaba's Tongyi MAI team just published benchmark numbers that make the case it is close to working. It also has not shipped the weights.

What actually happened

The Qwen-UI-Agent technical report went up on arXiv on July 30 under the title "Toward Next-Generation Real-World Centric Foundation GUI Agents." The pitch is a single model that runs on real devices across mobile, desktop, web, and search, mixing GUI interaction with direct command-line execution, and trained with online reinforcement learning on trajectories longer than 100 turns.

The reported scores: 82.1% on MobileWorld, 92.2% on MobileWorld-Real, 97.5% on AndroidDaily, 79.5% on OSWorld-Verified, 73.6% on WebArena, and 81.5% on ScreenSpot-Pro. On OSWorld-v2 it reports 40.0% partial progress. The paper positions those results as competitive with or ahead of Opus 4.8, Gemini 3.1 Pro, and GPT-5.6 Sol.

Now the part the benchmark table does not tell you. As of today, the Tongyi-MAI organization on Hugging Face publishes MAI-UI-2B and MAI-UI-8B — the previous generation, last updated in December and January. The MAI-UI repo carries an Apache 2.0 license. There is no Qwen-UI-Agent checkpoint. Coverage calling this a release is describing a paper.

Why a GUI agent changes what you can automate

Screen-driving is the workaround for vendors who will not give you an API. Every integration quote we have ever written has a line item for the one system that refuses to expose data. A model that can operate that system's UI turns a $30k custom integration into a scripted workflow. That is the actual prize here, not the benchmark.

A 2B model that clicks reliably beats a frontier model that clicks occasionally. MAI-UI-2B is small enough to run on hardware you already own. Screen automation is high-volume and repetitive — a hundred clicks per task, thousands of tasks per month. Per-token frontier pricing on that workload is a bad trade even when the accuracy is better.

Benchmarks are not your workflow. AndroidDaily at 97.5% measures consumer phone tasks. Your task is a 2009 inventory system with a nested frameset and a modal that appears only on Tuesdays. Take the top-line number as evidence the category is viable, not as a prediction about your screens.

Check what is downloadable before you plan around it. Apache 2.0 weights you can pull today are a dependency. A paper is a signal. Build with the first, plan for the second.

Sources: Qwen-UI-Agent technical report (arXiv), Qwen-UI-Agent project page, Tongyi-MAI on Hugging Face.