At the 2026 Snapdragon Summit, Qualcomm announced a collaboration with StepFun, Wulianghuo, and Longsys to optimize the StepEdge-Omni 30B-MoE model for the Snapdragon 8 Gen 6 mobile platform. The model utilizes a Mixture-of-Experts architecture, which activates only specific modules to reduce computational load and memory bandwidth requirements.
The optimization efforts, including inference engine improvements and heterogeneous scheduling, reduced memory usage by over 50% compared to standard industry solutions. This allows the 30-billion-parameter model to run locally on smartphones without cloud connectivity, handling tasks such as email drafting, itinerary planning, and calendar synchronization.
Performance tests indicate that the model achieves a prefill throughput exceeding 330 tokens per second and a decoding throughput of over 28 tokens per second. Industry experts note that deploying large MoE models on mobile devices could advance low-latency, privacy-focused AI agents, though scalability depends on hardware costs, power consumption, and reliability.
