Qualcomm and StepFun Run 30B MoE Model Directly on Smartphones

At the 2026 Snapdragon Summit in Hawaii, US, Qualcomm announced a partnership with StepFun, Wulianghuo and Longsys to bring the StepEdge-Omni 30B-MoE model to phones based on the 6th-gen Snapdragon 8 Ultimate mobile platform, with on-device adaptation, inference optimization and a live agent demo.

The model has 30 billion parameters and uses a mixture-of-experts design that activates only the needed parts to cut computing and memory bandwidth demands. With optimizations in inference engine, scheduling and storage, its memory needs drop by over 50% compared with common industry solutions, allowing stable phone operation.

Qualcomm says the solution runs fully on-device without cloud services, handling email understanding, trip planning, calendar sync, flight and hotel recommendations, itinerary sharing and email drafting. Tests show prefill throughput above 330 tokens per second and decode throughput above 28 tokens per second.

Industry observers say on-device large MoE models could boost low-latency, privacy-focused agent apps, but wider adoption still depends on device cost, power consumption, model reliability and user acceptance.