Qwen Intelligence Launches: Three SOTA Mobile AI Agents Arrive

The Alibaba Qwen team has unveiled Qwen Intelligence, built around the vision of making personal AI genuinely accessible. The release delivers three state-of-the-art AI agents designed for mobile scenarios, along with a comprehensive suite of mobile benchmarking tools. This marks a significant step in the evolution of large language models — moving beyond conversational Q&A toward real task execution on mobile devices.

Three Agents, Three Roles

Qwen Intelligence is not a single model. Instead, it's structured as three collaborative, specialized agents covering planning, execution, and creation — a design that reflects a growing trend in Agent development. Rather than asking one general-purpose model to do everything optimally, task decomposition and specialized agents working in concert tends to produce more practically useful results.

Mobile Planner Agent: Task Planning and Orchestration

The first agent, Mobile Planner Agent, handles the planning, decomposition, and scheduling of complex tasks. When a user presents a multi-step request that spans multiple apps, this agent breaks it into executable subtasks and coordinates their execution. According to official benchmarks, it ranks first on MobilePA-Bench, MobilePA-Bench Business, and Memory — indicating leading performance in both task planning and memory capabilities.

Mobile-Use Agent: Actually Getting Things Done

The second agent, Mobile-Use Agent, is the "hands-on executor." It follows an API-first, GUI-fallback strategy — calling structured interfaces whenever available, and only falling back to graphical UI interaction when no interface exists. This approach balances execution efficiency with broad compatibility.

The reported numbers are impressive: MobileWorld 82.1, MobileWorld-Real 92.2, AndroidDaily 97.2, with an end-to-end task success rate of 90%. That 90% figure deserves attention — completing cross-app operations on real mobile devices involves navigating UI changes, pop-ups, permission dialogs, and a host of unpredictable factors, making sustained high success rates genuinely difficult to achieve.

A note on API-first, GUI-fallback in mobile agents: API-first means the agent preferentially calls structured interfaces exposed by applications (such as Android's Accessibility Service or an app's own open APIs). These calls are fast, produce predictable states, and are unaffected by UI layout changes. GUI-fallback means reverting to human-like behavior — reading the screen and tapping buttons — when no usable interface exists, typically via screenshot recognition and coordinate-based tapping. The critical difference is stability: API calls yield deterministic outcomes, while GUI interactions are vulnerable to system pop-ups, dark mode switches, and manufacturer-customized UIs. This is why sustaining a high end-to-end success rate across long-tail applications is considered one of the most critical real-world dimensions for evaluating such agents.

Mobile Creative Agent: From a Single Sentence to a Finished Product

The third agent, Mobile Creative Agent, is positioned as a creation tool — capable of turning a one-sentence description into a ready-to-use output. The team particularly highlights speed: a single image generates in approximately 3 seconds, roughly 2× faster than leading comparable products. On mobile, where users are acutely sensitive to latency, generation speed often makes or breaks the experience.

On multi-agent architecture: Decomposing complex tasks across multiple specialized agents — a Multi-Agent System — has become the dominant paradigm for handling complex scenarios in recent years. The core logic is this: a single large model must constantly trade off between parameter capacity and inference speed when handling planning, tool use, and content generation simultaneously. Distributing different capabilities across purpose-built sub-models allows each component to be iterated and upgraded independently, and makes it easier to diagnose and fix failure cases. The tradeoff is increased system complexity — information passing and state synchronization between agents become new potential failure points. Qwen Intelligence's separation of planning, execution, and creation into three distinct agents is a concrete application of this architectural philosophy to the mobile domain.

An Open-Source Benchmark Suite

Alongside the three agents, the Qwen team has released a full set of mobile benchmarking tools covering multiple evaluation dimensions:

• MobilePA-Bench: Focused on planning capability assessment
• MobileWorld (open-sourced on GitHub): Cross-app execution capability
• MobileWorld-Real: Performance on real physical devices
• MobileWorld-Safety: Safety evaluation

Together, these benchmarks form a reasonably complete evaluation loop spanning planning, cross-app execution, real-device performance, and safety. Open benchmarks matter because they enable the broader community to compare different mobile agents on a common standard — rather than relying solely on vendor-reported numbers. That said, these benchmarks were proposed by the releasing party, and their authority will need to be validated through community adoption and scrutiny over time.

A note on mobile benchmark construction: Building mobile benchmarks is significantly harder than text or code evaluations. Real-device testing (like MobileWorld-Real) requires running full app interaction workflows on physical or emulated devices, and test environments are difficult to standardize — different device models, Android versions, and app versions can all affect results. This is a primary reason the industry has historically lacked widely accepted mobile agent benchmarks. It's also common for model developers to lead benchmark creation in the broader LLM space — similar patterns appear with OpenAI's SimpleQA and Google's BIG-Bench. The community typically needs months of replication experiments and adversarial testing before reaching consensus on a benchmark's objectivity and coverage.

Why Mobile Agents Matter

From an industry perspective, Qwen Intelligence continues the overarching 2025–2026 trajectory of large models moving "from answering to acting." Compared to desktop agents, mobile agents face a more fragmented application ecosystem, more restricted API permissions, and more complex interaction paradigms — making deployment harder. But they also address the high-frequency use cases that matter most to everyday users.

Packaging planning, execution, and creation into three specialized agents, backed by open-source benchmarks, is both a product launch and a strategic move to shape the narrative around how mobile agents are evaluated. For developers and researchers, the key questions worth tracking are: can these reported benchmark scores be reproduced in your own use cases, and how stable is the API-first, GUI-fallback execution strategy across long-tail applications in practice?

All data cited in this article comes from official release materials. Actual performance remains to be validated through independent third-party testing and real-world user feedback.