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

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

KEY FACTS

Qualcomm showcased an on-device MoE agent demo at the 9th China International Import Expo, developed in collaboration with Infris AI, Longsys, and StepFun.
The partners jointly optimized on-device MoE runtime efficiency across model, inference, storage, and chip platform layers.
Mainstream on-device models currently have 2 billion to 3 billion parameters, with complex task completion rates of roughly 60% to 80%.
MoE models partition parameters into multiple expert modules, invoking only a subset of them per inference.
On-device AI applications are currently mostly single-point features and are evolving toward agent experiences such as intent understanding and task orchestration.

KEY DATA

2 billion to 3 billionMainstream on-device model parameter scale

60% to 80%Complex task completion rate range

20MoE example expert module count

2 to 3Expert modules invoked per inference

PHYSIX OBSERVATION

On-device AI is moving from single-point features to system-level agents, and MoE is a realistic path around memory and power constraints. Qualcomm bringing in model, inference, and storage partners for full-chain optimization shows competition has shifted from chip specs to ecosystem collaboration. But a demo is just a demo: 2 billion to 3 billion parameter models still fall short on complex tasks, and real-world experience is the true test.

Source: QbitAI report