Qwen-UI-Agent Technical Report: Toward Next-Generation Real-World Centric Foundation GUI Agents

TL;DR

A foundation GUI agent from Tongyi MAI (Alibaba / Qwen AI Lab) that unifies mobile, computer, browser and DeepSearch use in one model, trained against sandbox environments plus a large-scale real-device mobile runtime. It reports 82.1% on MobileWorld — 14.6 points above Opus 4.8 — and 79.5% on OSWorld-Verified (source).

Authors & Org

Tongyi MAI, Alibaba / Qwen AI Lab. Report dated 2026-07-29, arXiv:2607.28227 (source).

The author list was not obtainable from this environment — arxiv.org was refused at CONNECT and no page of the report was read directly. Everything below comes from search-result extracts of the arXiv HTML and the Hugging Face paper page, and is recorded with that provenance (source).

Method

• A single model covering mobile, computer, browser and DeepSearch scenarios, rather than one specialist per surface
• Diverse sandbox environments combined with a large-scale real-device mobile runtime — the report's own emphasis is on real-device rather than simulator-only training
• A harness for proactive service initiation: it detects events such as a flight cancellation, reasons over the relevant affairs, and executes across platforms while maintaining task state
• The contribution is organised around what the report calls "five transitions" toward next-generation GUI agents; coverage spans real-device and benchmark-based mobile use, computer use, browser use, DeepSearch and GUI grounding

(source)

Results

Mobile use (source):

Benchmark	Qwen-UI-Agent
MobileWorld	82.1%
MobileWorld-Real	92.2%
AndroidDaily	97.5%
On MobileWorld the report states margins of +14.6 over Opus 4.8, **+12.0 over
GPT-5.6 Sol** and +8.9 over Seed 2.1 Pro percentage points.

Computer use (source):

Benchmark	Qwen-UI-Agent
OSWorld-Verified	79.5%
OSWorld-v2 (partial-progress)	40.0%
On OSWorld-Verified the report places it second overall, ahead of GPT-5.5,
Gemini 3.1 Pro and Seed 2.1 Pro — a weaker claim than the mobile one, and the report
makes it as such.

These are vendor-run figures with no published harness configuration, so per Eval Harness Configuration they are claims about a (model, harness) pair rather than properties of the model.

Significance

The mobile result is the part that is not incremental. A 14.6-point margin over Opus 4.8 on MobileWorld is not the shape of a frontier general model being edged out; it is a specialist beating generalists on their weakest surface. This wiki's Agents (LLM Agents) and Embodied Agents pages have tracked computer-use as the frontier labs' contested ground — this report argues mobile is separable from it, and that the separation favours whoever has real devices to train against.

The proactive service initiation harness is the other thing worth marking. Every GUI agent this wiki records is invoked; this one is described as detecting an event and initiating. That is a different trust surface, and the report does not appear to address it.

Open Questions

• Parameter count, architecture and license are unknown — not in any extract read. Search-result extracts say the models are on Hugging Face and ModelScope, but this was not verified against either hub and no license was obtained, so an open-weight release is a report claim here rather than a confirmed fact (source)
• No independent measurement. MobileWorld, MobileWorld-Real, AndroidDaily and OSWorld-v2 do not appear in either leaderboard snapshot this repo holds, so nothing here has been checked against a third party
• What "real-device mobile runtime" costs to reproduce, and whether the mobile margin survives a harness the authors did not build

Cite

Qwen-UI-Agent Technical Report: Toward Next-Generation Real-World Centric
Foundation GUI Agents. Tongyi MAI, Alibaba, 2026. arXiv:2607.28227.

• https://arxiv.org/abs/2607.28227
• https://huggingface.co/papers/2607.28227
• (snapshot)