Qwen-UI-Agent
Qwen-UI-Agent
Towards Next-Generation Real-World Centric Foundation GUI Agent
One agent that thinks, searches, and acts across mobile, desktop, and the web to complete real-world, long-horizon tasks.
MAI-UI Team ·
Alibaba Token Hub
PDFTechnical reportGitHub▶Watch demos
WHAT IT CAN DO
Built to complete real work across GUI interfaces.
01Mobile GUI Use02Computer Use03Computer GUI + CLI04Browser Use05Deep Research + GUI06Proactive Service
9:41
Set route & timeFlight MU5103 · 07:30
Current locationAirport Terminal 3
Tomorrow · 06:30→
9:41
Ride options · Budget ≤ CNY 100
Standard · CNY 868 min pickup · fastest in budget_✓_
Comfort · CNY 9912 min pickup · more space
Premium · CNY 1285 min pickup · over budget
→
9:41
✓Ready for approval
Tomorrow 06:30 · Airport T3
Standard · CNY 86 · 8 minApprove CNY 86_No booking before approval_
MOBILE GUI USE
CAPABILITY 01
Mobile GUI Use
Optimized for everyday tasks in real-device mobile GUI use, the model navigates changing Android apps, accounts, content, and network conditions to search, compare, schedule, shop, and coordinate reliably.
←01 / 06→
PERFORMANCE
End-to-End GUI Tasks
Mobile GUI UseComputer UseBrowser Use
MOBILE GUI USE
GUI-Only Success rate (%)
#### MobileWorld
907050
Scale 50–90
82.1
Qwen-UI-AgentAlibaba
73.2
Seed 2.1 ProByteDance Seed
70.1
GPT-5.6 SolOpenAI
67.5
Claude Opus 4.8Anthropic
62.3
Qwen 3.7 PlusAlibaba
58.1
Gemini 3.1 ProGoogle
Qwen-UI-AgentClosed-source ModelOpen-weight Model
REAL-DEVICE MOBILE GUI USE
Success rate (%)
#### AndroidDaily
10087.575
Scale 75–100
97.5
Qwen-UI-AgentAlibaba
95.2
Seed 2.1 ProByteDance Seed
93.8
Gemini 3.1 ProGoogle
93.0
Claude Opus 4.8Anthropic
92.6
GPT-5.6 SolOpenAI
79.8
Qwen 3.7 PlusAlibaba
Qwen-UI-AgentClosed-source ModelOpen-weight Model
COMPUTER USE
Benchmark score (%)
#### OSWorld-Verified
9077.565
Scale 65–90
79.5
Qwen-UI-AgentAlibaba
83.4
Claude Opus 4.8Anthropic
78.8
Seed 2.1 ProByteDance Seed
78.7
GPT-5.5OpenAI
76.2
Gemini 3.1 ProGoogle
73.3
Qwen 3.7 PlusAlibaba
Qwen-UI-AgentClosed-source ModelOpen-weight Model
LONG-HORIZON COMPUTER USE
Partial progress (%)
#### OSWorld-v2
60300
Scale 0–60
40.0
Qwen-UI-AgentAlibaba
54.8
Claude Opus 4.8Anthropic
49.5
GPT-5.5OpenAI
22.3
MiniMax M3MiniMax
22.1
Kimi K2.6Moonshot AI
21.5
Qwen 3.7 PlusAlibaba
Qwen-UI-AgentClosed-source ModelOpen-weight Model
BROWSER USE
Success rate (%)
#### WebArena
806040
Scale 40–80
73.6
Qwen-UI-AgentAlibaba
71.9
Claude Opus 4.8Anthropic
69.5
GPT-5.5OpenAI
65.3
Gemini 3.1 ProGoogle
59.0
Qwen 3.7 PlusAlibaba
48.4
GUI-Owl-1.5Alibaba
Qwen-UI-AgentClosed-source ModelOpen-weight Model
GUI Grounding
GUI GROUNDING
Grounding score (%)
#### ScreenSpot-Pro · No zoom
807060
Scale 60–80
76.6
Qwen-UI-AgentAlibaba
72.9
GUI-Owl-1.5Alibaba
69.6
UI-Venus-1.5Ant Group
68.9†
Qwen 3.7 PlusAlibaba
67.9
MAI-UIAlibaba
65.3†
Seed 2.1 ProByteDance Seed
Qwen-UI-AgentClosed-source ModelOpen-weight Model
GUI GROUNDING
Grounding score (%)
#### ScreenSpot-Pro · Zoom-in
857565
Scale 65–85
81.5
Qwen-UI-AgentAlibaba
80.7†
Seed 2.1 ProByteDance Seed
80.3
GUI-Owl-1.5Alibaba
79.0†
Qwen 3.7 PlusAlibaba
74.8
UI-Venus-1.5Ant Group
73.5
MAI-UIAlibaba
Qwen-UI-AgentClosed-source ModelOpen-weight Model
GUI GROUNDING
Grounding score (%)
#### SS-V2
1009590
Scale 90–100
97.5
Qwen-UI-AgentAlibaba
96.6†
Seed 2.1 ProByteDance Seed
96.6†
Qwen 3.7 PlusAlibaba
96.5
MAI-UIAlibaba
96.2
UI-Venus-1.5Ant Group
95.3
GUI-Owl-1.5Alibaba
Qwen-UI-AgentClosed-source ModelOpen-weight Model
GUI GROUNDING
Grounding score (%)
#### MM-GUI-L2
1009080
Scale 80–100
92.6
Qwen-UI-AgentAlibaba
91.3
MAI-UIAlibaba
90.9†
Seed 2.1 ProByteDance Seed
90.5†
Qwen 3.7 PlusAlibaba
88.6
UI-Venus-1.5Ant Group
86.8
GUI-Owl-1.5Alibaba
Qwen-UI-AgentClosed-source ModelOpen-weight Model
GUI GROUNDING
Grounding score (%)
#### OSW-G-R
857565
Scale 65–85
78.5
Qwen-UI-AgentAlibaba
78.2†
Qwen 3.7 PlusAlibaba
78.0†
Seed 2.1 ProByteDance Seed
75.0
MAI-UIAlibaba
70.6
UI-Venus-1.5Ant Group
69.7
GUI-Owl-1.5Alibaba
Qwen-UI-AgentClosed-source ModelOpen-weight Model
GUI GROUNDING
Grounding score (%)
#### UI-Vision
806040
Scale 40–80
70.0
Qwen-UI-AgentAlibaba
68.0†
Qwen 3.7 PlusAlibaba
62.0†
Seed 2.1 ProByteDance Seed
54.7
UI-Venus-1.5Ant Group
47.1
MAI-UIAlibaba
Qwen-UI-AgentClosed-source ModelOpen-weight Model
†Author-reproduced result: the baseline was independently evaluated in the authors’ environment rather than copied from the model provider’s report.
Broader Capabilities
Real-world tasks demand more than interface interaction—they also require knowledge, multimodal reasoning, instruction following, and tool use. Qwen-UI-Agent gains strong GUI capabilities without becoming a narrow GUI-only model, preserving the base model’s general reasoning and agentic strengths for broader tasks.
01
Foundational reasoning
Multimodal understanding, knowledge, mathematics, and instruction following.
Swipe horizontally to compare all models →
| Benchmark | Qwen-UI-Agent | Qwen3.5-27B | UI-Venus 30B-A3B | GUI-Owl 32B | OpenCUA-72B | | ---------------------------- | ------------- | ----------- | ---------------- | ----------- | ----------- | | MMMU-Pro | 72.4 | 73.5 | 32.4 | 39.5 | 31.0 | | RealWorldQA | 83.1 | 83.1 | 75.3 | 76.7 | 66.4 | | CharXiv-RQ | 77.7 | 76.8 | 44.7 | 50.9 | 39.6 | | MathVision | 82.8 | 82.0 | 36.8 | 50.6 | 26.6 | | AI2D\_TEST | 91.1 | 91.9 | 84.3 | 84.8 | 78.9 | | MMLU-Pro | 86.5 | 86.0 | 65.6 | 73.9 | 58.8 | | IFEval (prompt-level strict) | 90.2 | 90.4 | 81.3 | 84.5 | 70.6 |
02
Broader agentic work
Tool use, terminal tasks, multi-turn service work, coding workflows, and deep-research retrieval on BrowseComp (BC) and BrowseComp-ZH (BC-ZH).
Swipe horizontally to compare all models →
| Benchmark | Qwen-UI-Agent | Qwen3.5-27B | UI-Venus 30B-A3B | GUI-Owl 32B | OpenCUA-72B | | -------------------------- | ------------- | ----------- | ---------------- | ----------- | ----------- | | Tau2-Bench | 89.9 | 89.2 | 22.7 | 6.1 | 14.4 | | Terminal-Bench 2.0 · Avg 5 | 50.1 | 41.1 | 3.2 | 0.0 | 9.0 | | Claw-Eval · Avg 3 | 73.5 | 66.9 | 30.6 | 29.6 | 26.4 | | Claw-Eval · Pass@3 | 51.8 | 41.2 | 5.5 | 5.5 | 0.5 | | BFCL-v4 | 74.2 | 71.3 | 19.8 | 32.7 | 28.3 | | SkillsBench · Avg 5 | 28.0 | 24.9 | 0.5 | 0.3 | 0.0 | | QwenClawBench · Avg 3 | 44.2 | 48.5 | 6.4 | 5.1 | 11.4 | | BrowseComp (BC) | 64.1 | 61.0 | — | — | — | | BrowseComp-ZH (BC-ZH) | 75.0 | 62.1 | — | — | — |
EVALUATION NOTE
All scores were independently reproduced in the authors' evaluation environment. Some harness, judge, simulator, runtime, or task-subset settings differ from official evaluations and are documented in the technical report.
DEMOS
Choose a demo domain 05 domains
01Real-device Mobile UseLive apps & GUI02Computer UseDesktop workflows03Cross-device GUI UsePhone ↔ computer04Mobile Use + Deep ResearchSearch, verify & act05Proactive ServiceMonitor, recommend & act
Available workflowsReal-device Mobile Use
01Recipe research + e-shoppingI’m planning to make “passion-fruit sour-soup beef” tonight. Search Douyin for the most-saved photo-and-text post, save it, and remember the ingredients I need to prepare. Then, in the Hema app, purchase all the ingredients mentioned in the post—excluding seasonings—select delivery for 18:45 today, and place the order._02Hotel search + map verification_I’m visiting Hangzhou next Friday. On Trip.com, find hotels near West Lake, sort by guest rating, exclude homestays, and review the top three. Then use Amap to compare the walking time from each hotel to Broken Bridge. Book the hotel with the shortest walk, and tell me its location, price, and walking time to Broken Bridge._03Local discovery + review synthesis_I’m meeting a friend for coffee at Tianmuli today. Use Amap to find its full address. Open Dianping and find the most popular café within 1 km, and note its name. Then use Xiaohongshu to summarize the first five posts and tell me which coffee people recommend at that café._04Train planning + meeting scheduling_Next Wednesday I’m returning from Beijing. Check 12306 for the arrival time of the earliest high-speed train to Hangzhou West. Then find how long the metro trip from Hangzhou West to Alibaba Xixi Campus takes and calculate when I can reach the office. Finally, schedule a one-hour DingTalk meeting at the next full hour after arrival, titled “Post-Trip Project Sync,” with Zhang San and me, and set a five-minute reminder._05Housing search + commute comparison_Next month I’m moving near Future Sci-Tech City in Hangzhou. On Beike, search Yuhang Future Sci-Tech City for one-bedroom apartments around CNY 3,000, sort by distance, and note the community name and monthly rent of the top three. Then use Amap to check the metro commute time from each community to Alibaba Xixi Campus. Report each community, rent, commute time, and which has the lowest combined cost: monthly rent + commute time × CNY 2 per minute.
Demo case 01 / 05
Mobile GUI Use
Recipe research + e-shopping
Task instruction (translated from Chinese)
I’m planning to make “passion-fruit sour-soup beef” tonight. Search Douyin for the most-saved photo-and-text post, save it, and remember the ingredients I need to prepare. Then, in the Hema app, purchase all the ingredients mentioned in the post—excluding seasonings—select delivery for 18:45 today, and place the order.
Complete everyday workflows across changing apps on physical mobile devices.
The agent first extracts a recipe and its ingredients from Douyin, then carries that information into Hema to complete a time-constrained grocery order.
CITATION
BIBTEX · TEMPLATE□Copy template
@misc{qwenuiagent2026,
title = {Qwen-UI-Agent Technical Report:
Toward Next-Generation Real-World-Centric
Foundation GUI Agents},
author = {MAI-UI Team},
year = {2026},
note = {Alibaba Token Hub}
}