Disclosed is a method and device for processing mirror images of an AI platform, including: scanning a file used for building a mirror image to detect whether a format and syntax of the file are correct, and building the mirror image according to the file if the format and syntax are correct; selecting and installing a component required for mirror image training; adjusting queue positions of mirror images to be imported according to priorities of the mirror images, and importing the mirror images according to the queue positions; classifying and recommending mirror images, and performing a model training according to the mirror images selected; and in response to detecting that a storage of mirror images in a local disk space reaches a preset cleaning threshold, screening the mirror images, and cleaning mirror images screened out.