Alibaba launches Qwen Intelligence with mobile agents for cross-app tasks and agentic smartphones
Alibaba has launched Qwen Intelligence, a full-stack mobile AI platform designed to let smartphone manufacturers build agents that can plan complex tasks, operate across applications, use APIs and graphical interfaces, retain context, and continue workflows that extend beyond a single assistant response.
Introduced at the 2026 Apsara Conference, Qwen Intelligence combines mobile-optimized Qwen models with an agent harness, tool governance, device-cloud coordination and three initial production agents: Mobile Planner Agent, Mobile-Use Agent and Mobile Creative Agent.
HONOR is the first smartphone manufacturer adopting the platform. Alibaba and HONOR are co-developing domain-specific models and agent capabilities for MagicOS, with the upcoming HONOR Magic9 Series and HONOR Robot Phone among the first devices incorporating Qwen Intelligence.
The product is built around a shift from smartphone assistants that answer individual requests toward an execution layer that interprets an objective, decomposes it, selects tools, moves across apps and adapts when the state of the task changes.
........
QWEN INTELLIGENCE | ROLE |
Mobile Planner Agent | Planning, task decomposition, tool orchestration and dynamic replanning |
Mobile-Use Agent | Cross-app execution through APIs with GUI fallback |
Mobile Creative Agent | Mobile image generation and editing |
Foundation layer | Mobile-optimized Qwen models |
Platform layer | Harness, deployment, tools, evaluation and device-cloud coordination |
Application layer | Ready-to-use and industry-specific agents |
First smartphone partner | HONOR |
Deployment model | Full stack or individual modules |
........
··········
MOBILE PLANNER AGENT BREAKS A USER OBJECTIVE INTO A COMPLETE WORKFLOW
The Mobile Planner Agent acts as the coordination layer of Qwen Intelligence. Instead of converting a request into one immediate action, it can decompose a larger objective into subtasks, determine which tools or specialized agents are required, manage dependencies and replan when new information appears.
Alibaba uses the example of arranging a business trip to Shanghai. The planner can turn that instruction into connected operations involving flights, hotels, calendar entries and itinerary planning. If a selected flight is later canceled, the system is designed to use stored task state and available services to modify the workflow rather than forcing the user to reconstruct the original request.
That planning layer is central to smartphone automation because many useful tasks naturally cross application boundaries. Travel, shopping, communication, scheduling and payments often require several apps before the objective is actually complete, so Qwen Intelligence treats those operations as one workflow rather than a sequence of unrelated chats.
··········
MOBILE-USE AGENT COMBINES APIS WITH GUI CONTROL
The Mobile-Use Agent handles execution inside applications with a hybrid approach. Its preferred route is API-first execution: when an app exposes a reliable structured interface, the agent can call it directly instead of reproducing a sequence of taps and screen interactions.
When no suitable API exists, the system can fall back to graphical user interface control, interpreting the screen and operating the application more like a human user. The result is broader coverage without forcing every app developer to expose a dedicated agent API before the system can act.
........
EXECUTION METHOD | ADVANTAGE | LIMITATION |
API execution | Fast, structured and easier to verify | Requires supported integration |
GUI execution | Can operate apps without dedicated agent APIs | More sensitive to interface changes |
Hybrid execution | Selects the appropriate route during one workflow | Requires orchestration across both environments |
........
Pure GUI agents can potentially interact with almost any application but must continually interpret screen state, locate controls and recover from interface changes. Pure API agents are usually more deterministic but cannot act where integrations do not exist. Qwen Intelligence uses structured tools when available and reserves GUI navigation for the gaps.
Because mobile agents can interact with communications, calendars, files, accounts and payments, permission handling and tool governance become part of successful execution rather than an afterthought. The platform therefore places security boundaries around which resources an agent can access and how it can act on them.
··········
MOBILEWORLD TESTS THE TYPE OF CROSS-APP WORK THESE AGENTS NEED TO COMPLETE
Alibaba's mobile-agent work is accompanied by MobileWorld, an evaluation environment built around real execution rather than static question answering. It contains 201 tasks spanning 20 mobile applications and deliberately emphasizes long workflows and movement between apps.
The benchmark's tasks require an average of 27.8 completion steps, compared with 14.3 in AndroidWorld, while 62.2% of MobileWorld tasks involve multiple applications versus 9.5% in AndroidWorld. It also includes workflows where the agent must clarify ambiguous instructions with the user and tasks that combine graphical navigation with external tools through Model Context Protocol integration.
........
MOBILEWORLD CHARACTERISTIC | VALUE |
Tasks | 201 |
Applications | 20 |
Average task length | 27.8 steps |
Cross-app tasks | 62.2% |
Agent-user interaction tasks | 22.4% |
MCP-augmented tasks | 19.9% |
........
Those characteristics are closer to the operating conditions of a useful smartphone agent than a benchmark based on a single isolated interaction. A system that can press the correct button once is very different from one that preserves state across dozens of actions, several applications and an unexpected change midway through execution.
··········
MOBILEPA-BENCH EVALUATES PLANNING, MEMORY, TOOLS AND SUB-AGENTS
MobilePA-Bench focuses on the planner responsible for deciding what the mobile agent should do. It contains 1,705 tasks across 13 functional domains and 212 realistic mobile tools running inside a stateful executable environment.
Instead of scoring only a final response, the benchmark can inspect actual tool execution, resulting application state and the agent's behavior throughout the workflow. It covers basic tool execution, collaboration with specialized sub-agents, use of stored memory and invocation of preconfigured skills.
........
MOBILEPA-BENCH | SCALE |
Total tasks | 1,705 |
Mobile tools | 212 |
Functional domains | 13 |
Basic tool-use tasks | 1,040 |
Memory tasks | 376 |
Skill-use tasks | 200 |
Sub-agent tasks | 89 |
........
The environment also exposes agents to incorrect tool ordering, permission restrictions, missing information and runtime errors. Success therefore depends on executing the correct sequence, preserving state and recovering from failures rather than merely predicting what the next action should be.
··········
ALIBABA REPORTS STRONG RESULTS FOR ITS MOBILE-USE AGENT
Alibaba reports that the Mobile-Use Agent reaches 82.1 on MobileWorld, 92.2 on MobileWorld-Real and 97.2 on AndroidDaily, together with approximately 90% end-to-end task success in its published mobile-agent results.
These are vendor-reported evaluation results and should be interpreted within the specific harnesses, model configurations and test environments used for each benchmark rather than as directly interchangeable measures. They nevertheless show the metric Alibaba is optimizing: whether the agent actually completes a mobile workflow rather than whether the underlying model produces a strong textual answer.
Alibaba's evaluation stack spans MobilePA-Bench, MobileWorld, MobileWorld-Real and MobileWorld-Safety, covering planning, cross-application execution, real-device operation and safety behavior.
··········
HONOR IS THE FIRST HARDWARE PARTNER
Qwen Intelligence is already being integrated into devices through Alibaba's partnership with HONOR. The companies are jointly developing specialized models and mobile-agent functionality for MagicOS, with the HONOR Magic9 Series and HONOR Robot Phone among the first products expected to use the technology.
Alibaba reports that the resulting HONOR agent architecture can reach task accuracy of up to 91.8% and orchestrate workflows extending beyond 100 individual steps. At that length, the engineering problem is state preservation: the agent must retain the original objective, track completed operations, preserve outputs needed later, avoid repeating irreversible actions and respond correctly when an application returns an unexpected result.
··········
THE PLATFORM USES A THREE-LAYER ARCHITECTURE
Alibaba organizes Qwen Intelligence into three layers so smartphone manufacturers can adopt the entire system or integrate only the components they need. At the foundation layer are Qwen models optimized specifically for mobile environments; the middle platform layer supplies deployment, harness customization, unified tool governance, automated evaluation and coordination between local devices and cloud inference; the application layer contains finished agent capabilities and vertical solutions.
That modularity matters because smartphone manufacturers already control deeply integrated services such as contacts, photos, notifications, device settings and account permissions. A manufacturer can therefore integrate a specific Qwen agent, use Alibaba's mobile models with its own interface, customize the harness around proprietary device services or deploy the complete stack.
........
ARCHITECTURE LAYER | FUNCTION |
Foundation models | Mobile perception, reasoning and generation |
Agent platform | Harness, tools, deployment, evaluation and orchestration |
Premium agents | User-facing and vertical workflows |
........
··········
DEVICE-CLOUD COORDINATION IS BUILT INTO THE EXECUTION MODEL
Qwen Intelligence is designed around a combination of local and cloud computation rather than assuming every operation should run in the same environment. On-device execution can reduce latency, preserve privacy and retain access to device state, while cloud inference provides more compute for difficult planning, multimodal reasoning or long-running tasks.
The platform can route stages of a workflow according to performance, latency and cost requirements. Sending every action to a large cloud model can create excessive latency and usage cost, while forcing every component on-device introduces memory, thermal, power and model-size constraints. The practical mobile-agent architecture is therefore likely to mix models and execution environments within the same task.
··········
MOBILE CREATIVE AGENT MOVES GENERATION DIRECTLY INTO PHONE WORKFLOWS
The third initial component, Mobile Creative Agent, handles image generation and editing. Alibaba describes it as a lightweight model optimized for smartphones and reports image creation in approximately three seconds under its published configuration.
Its role inside the agent platform is more important than the existence of another image generator. A planning agent can invoke creative generation as one stage of a broader workflow, then pass the result directly to another application without requiring the user to manually transfer files between separate AI tools. Generation becomes a callable capability inside mobile automation rather than a standalone destination.
··········
MEMORY AND PROACTIVE ASSISTANCE ARE NEXT IN THE STACK
Alibaba says additional Qwen Intelligence services are in development, including personalized memory, proactive assistance, expanded multimodal interaction and specialized vertical agents built with ecosystem partners.
Memory is particularly important on smartphones because the device holds persistent personal context. A capable agent could use recurring preferences, previous decisions, schedules, contacts and prior task history to eliminate repeated configuration, but the same capability increases the importance of access controls and auditability.
Proactive assistance raises the bar again because the agent may initiate a workflow in response to changing conditions instead of waiting for an explicit command. Permission architecture, human confirmation and reversible actions become progressively more important as execution shifts from responsive to proactive.
··········
QWEN INTELLIGENCE MOVES THE SMARTPHONE AGENT BELOW THE APP LAYER
Most AI assistants on smartphones have historically existed as applications or interface features sitting above the operating system. Qwen Intelligence is designed for deeper integration by device manufacturers, allowing its planner, execution agent, mobile models, tool-governance system and device-cloud architecture to operate across applications instead of remaining confined to one AI app.
That changes the unit of automation from an individual app interaction to a user objective spanning the device. Organizing a trip, for example, naturally involves search, booking, scheduling, navigation, communication and potentially payment. A conventional assistant helps with individual pieces; an agentic operating layer attempts to coordinate the complete chain.
The approach also gives smartphone manufacturers a strategic reason to build their own agent layer because they control the operating system, hardware capabilities, permission framework and access to device services required for reliable cross-app execution.
··········
AGENTIC SMARTPHONES WILL BE MEASURED BY COMPLETED TASKS, NOT CHAT QUALITY
Qwen Intelligence reflects a broader shift in mobile AI from conversational performance toward execution reliability. The useful production metrics therefore become task-completion rate, number of actions required, frequency of GUI fallback, recovery from app errors, user-confirmation rate, local-versus-cloud inference share and total workflow cost.
The benchmarks surrounding Qwen Intelligence already move in this direction by evaluating long-horizon execution, memory, permissions, cross-app navigation, tool use and real-device behavior. Alibaba's commercial stack connects those pieces: a planner decides what should happen, specialized agents perform the work, APIs and interfaces expose applications, and the platform coordinates execution across device and cloud.
If that architecture proves reliable at production scale, the smartphone's AI layer becomes less about answering the user and more about operating the device on the user's behalf.
·····
FOLLOW US FOR MORE.
·····
·····
DATA STUDIOS
·····
[datastudios.org]
