Alibaba released Qwen-UI-Agent on August 20, 2026, a foundation model for agents that look at a screen and operate it. A single model covers smartphones, desktop computers, web browsers, and deep search, and on mobile benchmarks run against real devices it pulls ahead of competitors by more than 10 points. On desktop tasks, however, rival models still come out on top in several categories, leaving a clear picture of where the model is strong and where it is not.

Mobile is where the gap is widest

The published evaluation results show the largest margins in the mobile domain. On MobileWorld, which measures whether a task can be completed from screen information alone, Qwen-UI-Agent scored 82.1 percent. Next came Seed 2.1 Pro from ByteDance Seed at 73.2 percent, GPT-5.6 Sol from OpenAI at 70.1 percent, and Claude Opus 4.8 from Anthropic at 67.5 percent. That is a 12.0 point lead over GPT-5.6 Sol and a 14.6 point lead over Claude Opus 4.8.

On MobileWorld-Real, which runs on physical handsets rather than simulators, the model reached 92.2 percent, ahead of Seed 2.1 Pro at 88.7 percent and Google's Gemini 3.1 Pro at 86.2 percent. On AndroidDaily, a collection of everyday operations, it hit 97.5 percent, leaving almost nothing on the table.

Phone screens change layout constantly as apps update, and what the agent sees also shifts with login state and network conditions. Handling that variability well appears to be what produces the gap.

Desktop results are mixed

Computer-use results are less uniform. OSWorld-Verified came in at 79.5 percent, above Seed 2.1 Pro at 78.8 percent and GPT-5.5 at 78.7 percent, but short of Claude Opus 4.8 at 83.4 percent. On OSWorld-v2, which measures how far an agent gets through long multi-step tasks, the score was 40.0 percent, well behind Claude Opus 4.8 at 54.8 percent and GPT-5.5 at 49.5 percent.

Browser work told a different story. WebArena put Qwen-UI-Agent in first place at 73.6 percent. For GUI grounding, which measures how precisely the model can point at an element on screen, ScreenSpot-Pro returned 81.5 percent with zoom-in and 76.6 percent without it.

The reading, then, is that mobile and browser work are strong while sustained desktop work still has room to improve. Some baseline numbers were re-measured in the development team's own environment, so the figures are best treated as indicative.

Trained on more than 100 physical devices

The performance rests on the training and evaluation setup. Development used over 100 physical handsets and more than 150 applications, wired together so that task construction, tracking, training, and evaluation run as one continuous loop. Alibaba also built MobileWorld-Real, a benchmark of more than 400 tasks across 100-plus applications, aimed at closing the gap between simulators and real conditions.

On the training side, the team applied online reinforcement learning to sequences longer than 100 steps, running roughly 10,000 concurrent environments so the model learns to carry long procedures through to the end without losing the thread.

Mixing screen actions and commands in one decision

Another distinguishing feature is that the model switches between clicking and typing on screen and issuing command-line operations within the same flow. It can emit several actions from a single decision, and roughly 40 percent of desktop tasks involved batched command execution.

Plenty of work, such as file operations and configuration changes, is faster with one typed line than with a trail of clicks. The idea is to give the model the same reflex a human has when switching to a terminal, and fewer steps means fewer places to fail.

Models tuned heavily for GUI work tend to lose general reasoning ability, but that seems to have been contained here. MMMU-Pro for multimodal reasoning came in at 72.4 and IFEval for instruction following at 90.2, roughly matching the base model. Terminal-Bench 2.0, which measures terminal work, rose to 50.1 from the base model's 41.1.

Stopping short of payments and deletions

The obvious worry with an agent that operates a screen on its own is whether it will buy or delete something without being asked. Alibaba says the model refuses illegal or high-risk tasks and pauses before critical operations such as payments, data deletion, and privacy authorization, notifying the user at that point.

The published demos follow the same pattern: a ride booking is presented with price and estimated time and then waits for approval, with the final confirmation left to the person. Where to draw the line on agent autonomy is something every vendor is still working out, and being explicit about where the agent stops suggests a design aimed at real use.

Summary

Qwen-UI-Agent posted a clearly shaped set of results: more than a 10 point lead over rivals on real-device mobile operation, but behind Claude Opus 4.8 and others on long-running desktop work. The main talking points are the training environment built on 100-plus physical devices and 150-plus applications, the combined use of screen actions and the command line, and the safety design that halts before payments and deletions. With the technical report and demos publicly available, it is worth a look as a way to gauge where GUI agents currently stand.