Qwen Intelligence: Alibaba's Agent Platform for Smartphones

Most talk about AI on phones has so far centred on the chatbot: a box you type into, an answer that comes back. That works for looking up facts. It is less useful for someone whose train has been cancelled and who now needs to rebook the journey, move a meeting and tell a friend they will be late.

That is the kind of problem Alibaba is aiming at with Qwen Intelligence. It is not a single app you download. It is a platform that smartphone makers can build into their own system assistants, so its AI functions sit inside the phone rather than on top of it. The stated goals are practical: replanning trips, adjusting appointments and creating images.

One request, three agents: how the work is split

Qwen Intelligence divides a task between three specialised agents rather than asking one model to do everything.

The first is the Mobile Planner Agent. It takes a request and breaks it down into individual steps. If circumstances change, for example when a travel connection shifts, it adjusts the plan instead of starting over. With the user's permission, it can also take saved preferences into account.

The second is the Mobile-Use Agent, which carries out those steps. Where an app offers a direct interface, the agent uses it, which cuts out detours in how the app is operated. Where no such interface exists, it falls back on the screen itself: it recognises buttons and taps or swipes the way a person would. For payments and other sensitive actions, a confirmation step is built in. That matters, because an agent that can press buttons on your behalf is also one that raises the question of where the security boundary sits.

The third is the Mobile Creative Agent, which handles images. From a short sentence, it can produce a motif for a greeting card or a background. On speed, Qwen's own materials do not agree. The announcement post cites three seconds per image. The product documentation puts simple jobs at around six seconds and more complex ones at roughly ten.

Measuring the agents: what the benchmarks show

Alibaba has published results across several tests, and on each of them Qwen comes out ahead, though the margins vary.

On MobilePA-Bench, a planning benchmark, the Qwen agent scores 77 per cent. That places it just ahead of GPT-6 Astra at 76 per cent, a gap of a single point.

The lead is wider on MobileWorld, a test built around longer tasks that span several apps. There, the Mobile-Use Agent reaches 82 per cent, against 73 per cent for the next-best competitor.

On MobileWorld-Real, which runs on actual smartphones rather than in a simulated setting, Qwen records 92 per cent. The runner-up reaches 89 per cent.

These figures describe completed test tasks. They show how often the agents finished what they were asked to do under benchmark conditions, which is not the same as how they will behave across the full range of apps and situations on a real person's phone.

Operating the phone, not just talking to it

The design choice worth noting is the fallback in the Mobile-Use Agent. An agent that relies only on official app interfaces is limited to whatever developers choose to expose. One that can read the screen and tap through it can, in principle, work with apps that were never built for AI at all. The trade-off is that screen-based operation is the slower, less direct route, which is why the agent uses a proper interface first when one is available.

This puts Qwen Intelligence alongside a wider move towards assistants that act rather than answer, such as Google's approach of letting Gemini phone businesses on a user's behalf. The difference here is the delivery model: Alibaba is supplying the layer to device makers instead of shipping it as its own consumer product.

Who can use it, and when

For now, Qwen Intelligence is aimed at manufacturers, not consumers. Phone makers can test the public beta through Qwen's website and through Alibaba Cloud. There is no freely available version for private users.

The first handset announced with the technology is the Honor Magic9, due to be presented on 28 September. That launch will be the first chance to see how the three agents perform once they leave the benchmark and reach a device in someone's pocket.

For readers following the Qwen family more broadly, the platform sits next to efforts that run Qwen models locally on consumer hardware for coding. Qwen Intelligence takes a different path: rather than handing developers a model, it offers phone makers a ready-made set of agents to build into their systems.

The key question is no longer whether a phone assistant can hold a conversation. It is whether it can plan a task, work through the apps needed to finish it and stop to ask before it spends your money. On Alibaba's own numbers, Qwen Intelligence does that more often than its rivals. The Honor Magic9 will show whether that holds up outside the test lab.