Alibaba launches Qwen Intelligence, an agentic smartphone platform debuting on HONOR

Alibaba launched Qwen Intelligence, an agent platform for smartphones built with HONOR. The stack has three layers: mobile-optimized Qwen models running on or near the device, an orchestration layer that breaks user requests into steps, and specialized task agents. Mobile Planner handles multi-step planning, Mobile-Use operates apps by navigating their interfaces, and Mobile Creative handles content generation.

Alibaba claims up to 91.8% task accuracy and the ability to carry out action sequences of more than 100 steps on HONOR Magic9 phones. That second number is the more interesting one. Phone agents typically fail on long tasks because a single misread screen early on derails everything that follows, the same failure mode NVIDIA researchers targeted this week with PivotOPD, a training method that teaches agents to recover from early mistakes. If Alibaba's 100-step claim holds up in real use, it would mean phone agents can complete genuinely complex jobs like booking travel across multiple apps.

Competitively, this is China's answer to Apple's redesigned Siri AI, which shipped in iOS 27 with multi-app action execution, and to Google's Gemini on Android. Chinese handset makers including HONOR, Xiaomi and Huawei have been racing to ship agent features, and Alibaba is positioning Qwen as the model layer they license. That ties into Alibaba's cloud ambitions; the company expects AI Cloud revenue to grow more than 50% year over year. Alibaba's Qwen 4 is reportedly in training, with a roadmap aiming toward 10-trillion-parameter MoE models.

Caveats: the accuracy figure is Alibaba's own, without a named benchmark, and the sources are secondary outlets. App developers may resist agents driving their interfaces, as they have in other markets. Watch for availability outside China, which other handset makers sign on, and how the platform handles payments and personal data when agents act inside banking and messaging apps.