On-device agents will need cloud-level quality and a way to carry user context across devices to be widely used, executives from Qualcomm and LiquidAI said at Snapdragon Summit. They cited a performance gap between local and cloud models and the need for agents to learn preferences separately on each device. LiquidAI proposed a two-layer approach combining a memory layer and an agent layer to transfer recorded context between devices.

On-device AI models, previously constrained by smartphone memory capacity, can grow to 30 billion parameters through data placement design, Qualcomm said. At Snapdragon Summit 2026, Qualcomm executive vice president Chris Patrick said more capable AI agents need access to larger models with stronger reasoning and richer context. Qualcomm unveiled the Snapdragon 8 Elite 6th Gen platform, expanding Hexagon NPU shared memory and adding an element accelerator supporting 32,000 tokens.