Physix Frontier · News Briefing Card (Leiphone · Nov 8, 2026)

Qualcomm demos on-device MoE agent at CIIE with four partners

KEY FACTS

Wan Weixing, Qualcomm's head of AI product technology for China, delivered a speech in Shanghai on November 7.
On-device AI applications are evolving from single-point features toward agent experiences such as intent understanding and task orchestration.
Mainstream on-device models range from 2 billion to 3 billion parameters, with complex task completion rates of about 60% to 80%.
At the ninth China International Import Expo, Qualcomm showcased an on-device MoE agent in collaboration with Infris AI, Longsys, and StepFun.
The partners respectively optimize model intelligence, expert prefetching, UFS-to-memory loading, and chip platform routing.

KEY DATA

2 billion to 3 billionMainstream on-device model parameter scale

60% to 80%Complex task completion rate range

20Number of expert modules in MoE example

2 to 3Number of expert modules invoked per task

PHYSIX OBSERVATION

The focus of on-device AI competition is shifting from 'can it run a model' to 'how complex a task can it handle.' The combination of MoE and storage prefetching shows that chip compute alone is no longer enough; models, inference, and storage must be optimized together. For users, the real signal is that on-device agents are entering the engineering deployment phase, but whether 2 billion to 3 billion parameters can support system-level operations remains to be seen.

Source: Leiphone report