Alibaba Qwen launches Qwen Intelligence mobile agent suite at Yunqi Conference

The release narrative for Qwen Intelligence doesn't start with model capabilities, but first gives a constraint: tasks on phones are stateful—tools have dependencies and permissions, every step changes data, users often don't state their preferences fully, and plans must be revised based on feedback during execution. From this, the publisher concludes that general-purpose conversational models can't cover this.

This explains why the solution isn't an all-in-one assistant but is split by responsibility into three deliverable capabilities: planning, execution, and creation. In other words, it reframes the problem of "one model doing everything" into an engineering problem of "who takes on state, permissions, and feedback." This judgment comes from the publisher's own account and is a source_claim, not an independently verified conclusion.

Mobile Planner Agent is positioned as the planning layer, responsible for breaking a single sentence into executable steps, orchestrating tools, and revising the plan midway. The publisher's example is that "a business trip to Shanghai on Monday" would be broken down into booking tickets, booking a hotel, writing a calendar entry, and arranging the itinerary, then rebooking after a flight cancellation.

Technically, it's described as a "model × Harness" closed loop: the model handles understanding, planning, and decision-making, while the Harness manages state, memory, and skills. Memory is divided into two layers: working memory follows the current task, and long-term memory distills experience into a user profile. The publisher claims this Agent ranks first on MobilePA-Bench, MobilePA-Bench Business, and Memory.

What's noteworthy is precisely the long-term memory layer. Some commentators have pointed out that the phrase "distills experience into a user profile" mixes two different things: preferences (he habitually books morning flights) and authorization (how much he can approve, which account he can act on). Getting the former wrong just requires an apology; getting the latter wrong means unauthorized execution. This is external commentary rather than a publisher statement, but the boundary issue it points to is indeed not separately broken out in the solution text.

Mobile-Use Agent is the execution layer, and its route is to use interfaces whenever possible: MCP, API, DeepLink, CLI; when no interface exists, it falls back to looking at the screen and tapping the UI. This priority ordering itself shows that the publisher treats GUI operation as a fallback rather than a first choice.

Sensitive actions have boundaries drawn: high-risk requests are refused, payments and deletions are handed to the user for confirmation, and actions like posting comments are handed back to the user. In other words, the closer an operation is to being irreversible or involving funds, the lower the degree of automation.

The execution-side figures given by the publisher include MobileWorld 82.1, MobileWorld-Real 92.2, AndroidDaily 97.2, and a 90% end-to-end success rate. These figures all come from the publisher's own statements, via secondhand paraphrase, with no third-party reproduction yet, and can only be treated as source_claim.

Mobile Creative Agent is the creation layer: the user speaks, and the planning model first distills it into a clear instruction, then produces a directly usable image—for example, generating a laid-out retro diary from a single sentence. The release notes state that distillation and reinforcement learning compressed generation from 100 steps to 8 steps, with the first image taking about 3 seconds—roughly twice as fast as major competitors. These speed and multiplier figures likewise belong to the publisher's own statements.

Released alongside the three Agents are four benchmarks, each corresponding to a question that was previously easy to handle vaguely: MobilePA-Bench measures stateful tool planning, covering tool invocation, memory, skills, and sub-Agent collaboration; MobileWorld measures long-horizon, cross-app execution in cloud environments, including clarifying with the user and MCP tools; MobileWorld-Real measures everyday tasks on real devices, with real accounts and networks; MobileWorld-Safety measures whether unsafe actions are taken in risk scenarios.

Releasing the evaluations together is equivalent to putting the standard for "what makes this good" on the table as well. But the benchmarks are proposed by the publisher at the same time, meaning the evaluation yardstick and the product come from the same party; until third parties reproduce these benchmarks, the rankings and scores are better read as claims to be tested rather than established facts.

Looking at the three layers separately, the planning layer takes on state and memory, the execution layer takes on permissions and irreversible actions, and the creation layer takes on instruction clarification. The weak link they all share is precisely long-term memory: it must remember user preferences while potentially touching the scope of user authorization.

The publisher's text folds both into the expression "user profile," without explaining whether preferences and authorization belong to different storage or different confirmation flows. External commentary has already pointed out this risk: getting a preference wrong can be apologized for, but getting authorization wrong means unauthorized execution. This is an inference based on the source, not a problem the publisher has acknowledged.

Therefore, what can currently be confirmed about Qwen Intelligence is the solution's structure and the publisher's self-reported capability figures; what cannot be confirmed is the stability of these capabilities under real accounts, real networks, and long-term use, as well as how the permission boundary of long-term memory is constrained. Although MobileWorld-Safety is among the four benchmarks, it measures whether unsafe actions are taken in risk scenarios, and does not directly answer how authorization memory is separated.