Running lightweight models directly in the browser places high demands on device computing power, typically requiring more than 8GB of memory with limited model parameter sizes; the local inference service approach offers greater flexibility but has a higher technical barrier.