Alibaba splits phone automation across three Qwen agents

Alibaba’s Tongyi Lab has released Qwen Intelligence, a three-agent stack for planning phone-based errands, operating apps and generating images. The lab also published MobileWorld, an open-source benchmark designed around longer workflows that cross app boundaries, require clarification and combine screen interaction with external tools.

Mobile agents turn natural-language requests into sequences of app actions. Qwen Intelligence divides that process among three components:

• Mobile Planner Agent: decomposes requests, coordinates subagents and tracks task state.
• Mobile Use Agent: operates the phone through structured app tools when available, with taps and swipes as a fallback. Alibaba publishes it as Qwen-UI-Agent.
• Mobile Creative Agent: generates an image from a single-sentence prompt in about three seconds. Alibaba claims roughly twice the speed of leading peers, although meaningful latency comparisons require matching hardware and generation settings.

Availability differs by component. MobileWorld’s code and evaluation assets are public, the planner has its own repository, and the other agent details appear on project pages. Deployment teams will need to verify model weights, licenses, supported apps and device requirements for each component. The performance figures below are vendor-reported and await independent reproduction.

MobileWorld stretches the workflow

MobileWorld contains 201 tasks across 20 applications. Its tasks average nearly twice as many steps as AndroidWorld, a commonly used Android-agent benchmark, and they involve cross-app workflows far more often.

Two task categories expand the evaluation beyond fixed instructions and screen navigation:

• Agent-user interaction tasks: 22.4% of the suite tests whether an agent asks useful questions when instructions are ambiguous. A GPT-4.1-based simulated user supplies the answers, so the interaction remains repeatable.
• MCP-augmented tasks: 19.9% requires agents to combine graphical interface control with external tools exposed through the Model Context Protocol, or MCP.

State checks make runs reproducible

MobileWorld runs inside Docker with a full Android emulator, self-hosted application back ends and deterministic verification. The evaluator checks database records, local storage and app callbacks to determine whether an agent reached the required end state.

Self-hosted services let the benchmark include backend-dependent applications such as messaging without relying on changing third-party APIs. End-state inspection also reduces the ambiguity of visual or manual grading, where a screen may appear correct even though the underlying transaction failed.

Qwen-UI-Agent mixes tools with touch

Alibaba reports that Qwen-UI-Agent scores 82.1 on MobileWorld, compared with 73.2 for Seed 2.1 Pro and 70.1 for GPT-5.6. The model reaches 92.2 on MobileWorld-Real and 97.5 on AndroidDaily.

Scores from separate benchmark variants are not directly interchangeable. Differences in task selection, environment and available tools can affect results, so the 10.1-point gap between MobileWorld and MobileWorld-Real cannot be assigned to API or MCP use without a controlled ablation.

Qwen-UI-Agent also retains capabilities outside phone control. It scores 72.4 on the multimodal MMMU-Pro benchmark, 82.8 on the visual-mathematics benchmark MathVision and 50.1 on Terminal-Bench 2.0. Alibaba reports Terminal-Bench scores of 3.2 for UI-Venus 30B and 0.0 for GUI-Owl 32B, two models specialized for graphical interfaces.

Desktop interaction remains less reliable. Qwen-UI-Agent records 40.0 partial progress on OSWorld-v2, behind Claude Opus 4.8 at 54.8. OSWorld evaluates computer use across longer desktop tasks, and the result points to continuing problems with precise interface grounding and extended action sequences.

The planner trains on execution traces

The Qwen-Planner-Agent uses a 27-billion-parameter model with a stateful harness that stores memory, manages reusable skills and dispatches subagents. On MobilePA-Bench, Alibaba reports an overall score of 77.05%, which is 9.83 percentage points above the team’s baseline.

Alibaba estimates an output-token cost of $2.41 per 1,000 tasks. That estimate excludes input tokens, device infrastructure and application hosting, all of which matter when calculating production cost.

The team calls its training method CARE, short for competence-adaptive planner training. It combines planning-focused supervised learning with reinforcement learning across mixed environments. CARE adjusts reward and advantage estimates according to the model’s current competence, giving less weight to already-mastered cases while scaling difficult failures to keep training stable.

Execution traces from deployed agents feed back into task construction, trajectory selection and planner training. The harness can use those traces to identify recurring failures, curate better examples and refine the skills available to later runs.

Recipes become multi-app test cases

Alibaba’s demonstrations illustrate the intended scope. In one task, the agent searches Douyin for the most-saved post about a recipe, extracts the ingredients, opens the Hema grocery app, excludes seasonings, adds the remaining items and schedules delivery for a specified time. Another task combines hotel search on Trip.com with walking-time comparisons in Amap.

These workflows test several capabilities that short, single-app tasks rarely cover:

• Maintaining state across multiple applications and long action sequences
• Applying constraints such as budgets, exclusions and delivery windows
• Requesting clarification before committing consequential actions
• Combining structured APIs or MCP tools with screen-based interaction
• Recovering when an app, tool call or intermediate step fails

APIs join the agent surface

MobileWorld gives developers a way to test architectures that combine planners, tool calls and graphical control under reproducible conditions. Its backend instrumentation also covers application categories that deterministic suites have often avoided because they require messaging, accounts or server-side state.

The design suggests several practical choices for mobile-agent systems:

• Expose stable application actions through APIs or MCP tools where possible.
• Retain graphical control for applications and actions without structured interfaces.
• Verify outcomes from application state, databases and callbacks rather than screenshots alone.
• Separate planning from execution so each layer can be evaluated and improved independently.
• Treat clarification and approval as explicit steps in workflows with ambiguity or financial consequences.

MobileWorld is available on GitHub, with links to its dataset and public leaderboard. The self-hosted applications and deterministic checks make it suitable for comparing agent revisions without score drift from changing external services.