Steven Hoi, Head of Qwen Mobile AI at Token Foundry and Vice President of Alibaba Token Hub (ATH), launches Qwen Intelligence at the Apsara Conference.
Alibaba has introduced Qwen Intelligence, a full-stack agentic AI platform designed to help smartphone manufacturers build devices capable of planning, executing, and adapting to complex tasks.
HONOR is the first smartphone maker to adopt the platform and is working with Alibaba to develop domain-specific models and solutions for its next-generation AI smartphones. HONOR’s Magic9 Series, scheduled to launch on September 28, and the HONOR Robot Phone are set to be among the first devices incorporating Qwen Intelligence.
Unveiled at the Apsara Conference in Hangzhou, Qwen Intelligence spans the AI stack, from mobile-optimized foundation models to ready-to-use agents. Manufacturers can adopt the complete platform or integrate individual components into their own systems.
Three agents for smarter smartphones
The first version includes three agents designed for different mobile AI tasks.
The Mobile Planner Agent functions as a task-planning layer, breaking complex requests into individual actions and coordinating tools. For example, a business-trip request could involve arranging flights and hotels, updating a calendar, and creating an itinerary. The system is also designed to adapt when circumstances change, such as a canceled flight.
The Mobile-Use Agent performs actions across applications using an API-first approach with a graphical user interface fallback. Alibaba says the architecture incorporates security boundaries and privacy protections for cross-app tasks.
The Mobile Creative Agent is a lightweight model optimized for smartphone-based image generation and editing.
Additional capabilities in development include personalized memory, proactive assistance, multimodal interaction, and specialized agents from ecosystem partners.
Three-layer architecture
Qwen Intelligence uses a three-layer architecture. At its foundation is a Qwen-based model optimized for mobile scenarios, designed to provide stronger on-device capabilities while balancing computing costs.
The middle platform layer supports customizable model deployment, tool governance, evaluation, and device-cloud coordination to balance performance, latency, and cost.
The top layer consists of specialized agents designed for specific real-world applications and industry requirements.
New benchmarks for agentic phones
Alibaba has also developed an open evaluation system to measure agent performance in real mobile scenarios. Its MobilePA-Bench covers more than 1,000 real-world tasks and over 200 commonly used mobile tools, testing an agent’s ability to plan and complete complex tasks.
Other benchmarks evaluate long-running, cross-application tasks and scenarios performed on cloud-based and physical devices, including safety and successful task execution.
Through its collaboration with HONOR, Alibaba is integrating Qwen Intelligence with HONOR’s MagicOS to develop what the companies describe as an agentic smartphone experience.
Alibaba and HONOR report that AI smartphones using the architecture can achieve up to 91.8% task accuracy and orchestrate workflows involving more than 100 steps, pointing to a shift from smartphones that primarily respond to commands toward devices capable of executing multi-step tasks.
