Alibaba has officially launched Qwen Intelligence, an end-to-end agentic AI solution designed to empower global smartphone manufacturers to embed autonomous capabilities directly into their devices. Unveiled at the annual Apsara Conference in Hangzhou, the platform spans everything from mobile-optimized foundation models to ready-to-deploy agents, providing hardware makers with the essential building blocks for next-generation intelligence.
Global smartphone manufacturer HONOR has emerged as the first partner to adopt Qwen Intelligence. The technology is integrated into HONOR’s MagicOS to power the new flagship HONOR Magic9 Series (launched September 28) and the HONOR Robot Phone. By shifting mobile devices from passive respondents to proactive service assistants, this joint architecture enables HONOR’s AI smartphones to achieve up to a 91.8% task accuracy rate while coordinating complex tasks spanning over 100 distinct steps.
Steven Hoi, Head of Qwen Mobile AI at Token Foundry and Vice President of Alibaba Token Hub (ATH), stated that agentic smartphones serve as the primary gateway connecting personal AI to the physical world. Qwen Intelligence combines mobile-optimized Qwen models with specialized control systems, allowing manufacturers to deliver tailored, proactive intelligence directly to end-users. The platform is accessible via its official website and through Alibaba Cloud services.
Three Ready-to-Use Agents for Everyday Operations
To help manufacturers build devices that better understand, plan, and execute tasks, the first phase of deployment introduces three core agents:
- Mobile Planner Agent: Serves as the cognitive engine for the phone, handling task decomposition, tool orchestration, and dynamic adaptation. For example, if tasked with “Organize my business trip to Shanghai this Monday,” it autonomously coordinates flight bookings, hotel reservations, calendar scheduling, and itinerary planning. If a flight is canceled, proactive memory systems immediately offer rebooking alternatives.
- Mobile-Use Agent: Executes cross-application actions through a hybrid mechanism prioritizing APIs while utilizing Graphical User Interface (GUI) interactions as a fallback, all within strict security and privacy boundaries.
- Mobile Creative Agent: A lightweight model tailored for on-device content creation, handling text generation and smooth image editing.
Future roadmap initiatives include personalized memory services with proactive assistance, multimodal interaction capabilities, and high-performance domain-specific agents built alongside ecosystem partners.
Three-Layer Architecture and Open Benchmarking Standards
The architecture of Qwen Intelligence provides hardware makers with modular flexibility across three levels:
- Foundation Layer: Houses mobile-tailored Qwen foundation models that deliver robust on-device performance at lower operational costs compared to generic alternatives.
- Middle-Platform Layer: A customizable, modular layer managing tool governance, control system adaptation, and automated evaluations, backed by device-cloud coordination to optimize performance, latency, and cost.
- Top Layer: Features premium, scenario-specific agents pre-configured into high-performance solutions for immediate practical deployment.
Addressing the industry-wide absence of systematic mobile agent evaluation, the Qwen Intelligence team introduced an open evaluation framework comprising four benchmark suites. Centered around MobilePA-Bench, the benchmark tests end-to-end planning and execution across more than 1,000 real-world tasks and over 200 mobile tools. Additional benchmarks evaluate multi-step, cross-app execution across cloud environments and real devices, confirming Qwen Intelligence’s leading industry performance.
