The 'thin client + cloud large model' architecture can reduce per-unit hardware costs and support rapid model iteration, but the cost is that every sentence of conversation must leave the device.