Alibaba’s Qwen Intelligence Wants to Give Your Phone a Brain and a Body

A phone assistant that understands everything but cannot book a flight is not an assistant. It is a chat window with a microphone.

That paradox has trapped phone AI for years. Models keep getting smarter. The experience barely changes. Ask an assistant to book a flight, and it opens the airline app, then stops. It cannot switch apps, fill a form, or tap pay. The model is not the problem. The problem sits outside it: app walls, system permissions, on-device compute, memory, and safety rules. Remove one piece and the chain from understanding to action breaks.

On September 22, 2026, at the Apsara Conference, Alibaba announced Qwen Intelligence, or QI. It is a full-stack AI solution for phone makers. Honor’s Magic9 series will be the first phone to ship with it on September 28. Honor’s Robot Phone will also support it. Alibaba will not build phones. It sells QI to the companies that do.

The pitch is simple. Other companies sell the brain. QI trains the brain and the body together.

Three Agents, One Loop

A phone Agent needs to plan a task, execute it, and create content along the way. QI’s first three Agents cover those jobs.

Mobile Planner Agent is the brain. It breaks a request into steps, picks tools, and replans when something changes. Say, “Book me a trip to Shanghai on Monday.” It splits the work into a flight, a hotel, a calendar entry, and a schedule. If the flight is canceled, it finds rebooking options and moves the meetings.

Two technologies sit under it. The first is model and Harness co-evolution. The model handles understanding, planning, and decisions. The Harness manages state, memory, and skills. They run in real environments, feed each other results, and improve together. The brain learns what the body can do. The body learns what the brain wants.

The second is dual-layer memory. Working memory follows the current task and disappears when the task ends. Long-term memory turns repeated behavior into a user profile, then uses that profile in the next plan. You fly to Shanghai often. You prefer one hotel. Your trips usually last two days. The user does not enter that. The system learns it.

Mobile-Use Agent is the hands. This is where most phone Agents fail. A phone has dozens or hundreds of apps, each one a closed island. An Agent that crosses apps has two paths. Where an API exists, it calls the API through MCP, DeepLink, or CLI. That path is short, fast, and stable. Where no API exists, it uses visual understanding to tap the screen like a person. That path is slower, but it covers the long tail.

QI uses both. API first. GUI when needed. Pure API misses too much. Pure GUI is too slow and too fragile. The hybrid is the only path that covers real use.

Mobile Creative Agent handles image and video creation. It turns spoken instructions into clear commands. Distillation and reinforcement learning cut the generation process from 100 steps to 8. The first image takes about 3 seconds.

The Numbers on Real Phones

QI is not a slide deck. Honor and Alibaba built a domain solution on Qwen Intelligence and Honor MagicOS. The joint system hit 91.8% overall task accuracy. GUI actions took 3.6 seconds. It handled long tasks with more than 100 steps. End-to-end service completion reached 90%.

On public benchmarks, Mobile-Use Agent ranked near the top in six evaluations, including MobileWorld, MobileWorld-Real, and AndroidDaily. It scored 82.1% on MobileWorld, 12 to 14 points ahead of GPT-5.6 Sol and Claude Opus 4.8.

Benchmarks matter more than stage lights.

Three Paths, and Alibaba’s Fourth

Apple keeps Siri inside its own hardware and software. The experience is consistent and secure, but it only runs on Apple devices. In early 2026, Apple said its next foundation model would use Gemini, but its focus stayed on AI controlling Apple software. It has not shown interest in letting AI run every app.

Huawei integrates vertically. Under HarmonyOS NEXT, Xiaoyi has deeper system access and stronger multimodal abilities. Its app market remains separate.

Google puts Gemini inside Android. It reaches many devices, but it cannot guarantee one experience or one business model.

Alibaba chose a fourth path. It does not bind itself to one phone maker. It does not own a phone entrance. It sits between the chip and the operating system as infrastructure.

That path is light on assets. No factories. No supply chain. It is also exposed. If phone makers pick another vendor or refuse deep integration, the money gets harder.

Alibaba is betting on standardization. If several phone makers connect to QI, Alibaba’s Agent layer reaches more devices than Apple’s and stays more controllable than Google’s.

That explains the QI launch roundtable. The head of AI at OPPO Research Institute, the AI vice president of vivo, and a Qualcomm global vice president joined. It was not polite networking. It was a display of partner reach.

Edge and Cloud Split the Work

Most users will never notice this. The industry should.

A long task changes its compute needs as it moves. Planning needs a large reasoning model. Execution may need only a small on-device model. Image creation needs cloud compute. Network speed, memory, power, and load change every second.

Older systems fix the split. Everything on the phone or everything in the cloud. QI moves the split with the task. A 4B-parameter model handles simple work on the device with no connection. Harder reasoning and generation go to Qwen in the cloud.

That puts new demands on chips and operating systems. At the 2026 Snapdragon Summit, Qualcomm announced its sixth-generation Snapdragon 8 Elite with Oryon CPU, Adreno GPU, and Hexagon NPU. The chip is built to run on-device Agents. Chipmakers and model makers now work on the same problem.

Safety Has Three Layers

An Agent that can operate your phone needs limits. QI uses three.

The first layer refuses illegal or high-risk requests.

The second layer sends key decisions back to the user. For money, data deletion, or privacy access, the Agent fills the form and prepares the plan. You tap the final button.

The third layer follows platform rules. For comments and messages that platforms do not allow an Agent to send, the Agent writes the content. You press submit.

The system does not block everything. It also does not do everything it can. It has boundaries.

QI also released four benchmark sets: MobilePA-Bench for planning, MobileWorld for task execution, MobileWorld-Real for real devices across 104 apps, and MobileWorld-Safety for safe action. The phone industry has lacked a standard way to measure real Agent work. QI handed over both the ruler and the test.

What Can Go Wrong

The architecture is clear. The rollout is not.

Third-party cooperation is the first wall. The API-first plan depends on apps opening interfaces. If WeChat, Alipay, and banks keep blocking cross-app Agents, the GUI fallback gets slower and less reliable. ByteDance’s Doubao phone assistant hit that wall at the end of 2025 when WeChat and Alipay blocked it. QI has not shown a way around it yet.

Team stability is the second. In March 2026, Lin Junyang, technical lead of the Qwen large model, left. Yu Bowen, head of Qwen post-training, left in the same period. Hui Binyuan, head of Qwen Code, joined Meta in January. Alibaba CEO Eddie Wu said in an internal email that the company will keep its open-source model strategy and increase AI R&D spending. It also formed a foundation model support group. The effect of those departures on model iteration needs watching.

The business model is the third. QI is licensed. Phone makers’ willingness to pay, and users’ willingness to pay for proactive service, are still untested.

Conclusion

Phones hold 90% of users’ time and 90% of their context. They are the first real-world entrance for personal intelligence.

The last decade belonged to super apps. Shopping on Taobao. Chatting on WeChat. Searching on Baidu. The Agent era changes the object of interaction. You do not open an app, find the right button, and tap through steps. You tell the phone what you want. The Agent crosses apps and does it.

IDC says the AI industry is moving from generative AI into agentic AI in 2026. Qualcomm CEO Cristiano Amon said at the Snapdragon Summit that Agents are becoming the main source of AI token demand and will shape AI architecture and economics.

Honor Magic9 sales will show whether QI is the turning point. The technical architecture and industry logic point in the right direction.