This application discloses a video classification model construction method and apparatus, a video classification method and apparatus, a device, and a medium. The method is performed by a computer device, and may include: obtaining training dataset, the training dataset including a plurality of video frames; inputting the training dataset into a to-be-trained classification model, the to-be-trained classification model including a plurality of convolutional layer groups and one output layer group, some of the convolutional layer groups including a non-local module; extracting long-distance dependency features of the plurality of video frames by using the non-local module, the long-distance dependency features being used to represent association relationships between spatial positions of the plurality of video frames; outputting the long-distance dependency features to the output layer group, to generate predicted classification results of the plurality of video frames; and training the to-be-trained classification model through a prediction loss function until a parameter in the to-be-trained classification model converges, to obtain a video classification model.