Alibaba has no interest in building smartphones. What it wants is to sit inside everyone else’s. At the 2026 Apsara Conference, Alibaba announced Qwen Intelligence, a modular phone-agent platform designed for OEM integration, with Honor’s upcoming Magic9 series named as the first formal adopter. This is a software and model play, full stop. The framing matters because it defines who Alibaba is competing with and what it’s actually selling.
The stack is built on the Qwen model base and ships three distinct agent types. Mobile Planner Agent handles complex task decomposition, like converting a business-travel request into flight bookings, hotel reservations, calendar updates, and real-time replanning when a flight is cancelled. It uses a dual-layer memory system combining working memory with long-term context. Mobile-Use Agent handles on-device execution, preferring API-first paths through MCP, DeepLink, or CLI, and falling back to GUI vision when those interfaces aren’t available. It also includes layered safety gates that surface high-risk actions, like deleting data or moving money, back to the user rather than executing autonomously. Mobile Creative Agent takes spoken creative briefs and produces layout-aware images. Alibaba claims the generation pipeline has been compressed from roughly 100 steps to 8, with first-image latency around three seconds.
Binding all of this together is what Alibaba calls the OS Harness, a layer that co-evolves with both the model and the device OS to handle power, memory, and latency constraints at the hardware level. Honor product-line president Fang Fei confirmed that the Magic9 series, expected around September 28, and Honor’s robot phone will be among the first devices running Qwen Intelligence, with the stack co-tuned against MagicOS. Alibaba vice president Steven Hoi was direct about the company’s position: Alibaba will not manufacture phone hardware. It’s a model-and-Harness supplier. OEMs keep control of the experience and user data.
The competitive context here is important. ByteDance has been pushing its own Doubao Phone Agent, and both companies are targeting the same wedge: becoming the default AI agent infrastructure for Android device makers in China. Meanwhile, Apple is still navigating its own on-device intelligence rollout with Apple Intelligence, and Google’s Gemini is embedded across Pixel and third-party Android devices globally. Alibaba’s approach is different in that it’s explicitly modular and OEM-agnostic, which gives it a broader potential distribution surface if other hardware partners follow Honor’s lead.
The release also ships four open evaluation sets covering planning, cross-app execution, on-device tasks, and safety scenarios. Partner-stated benchmarks cited in secondary Chinese coverage include roughly 91.8 percent complex-task accuracy and around 3.6-second GUI operation latency. Those numbers come from joint-solution claims rather than independent audits, so they should be treated as directional rather than definitive.
What Qwen Intelligence signals, more than any specific benchmark, is that the battle for the AI layer on mobile devices is moving fast in China. Alibaba is betting that OEMs want a capable, modular stack they can brand and differentiate rather than shipping a generic assistant. That’s a reasonable bet. The question is whether the Honor partnership stays exclusive or whether Qwen Intelligence starts showing up on devices from Xiaomi, Oppo, or Vivo in the next product cycle. That’s when this stack becomes something worth watching at a global scale.
