This application relates to a model training method. The method includes retrieving a current group of training samples, the training samples being based on a training set; obtaining first sample features of training samples in the current group of training samples based on a to-be-trained model; and obtaining, center features respectively corresponding to the training samples; obtaining feature distribution parameters corresponding to the training samples, the feature distribution parameter corresponding to each training sample being obtained by collecting statistics on second sample features of training samples in the training set that belong to the same classification category, and the second sample feature of each training sample being generated by a trained model; obtaining, based on the center features and the feature distribution parameters, a comprehensive loss parameter corresponding to the current group of training samples; and adjusting model parameters of the to-be-trained model based on the comprehensive loss parameter.