Video classification is a key task for organizing the massive volume of video data generated today. However, few-shot video classification, where only a small number of labeled examples are available, remains highly challenging. Most recent methods rely mainly on internal similarity or contrastive scores within a single network, which limits their ability to learn robust and generalizable representations. We propose a mutual-learning–based few-shot video classification approach, in which multiple networks are trained on the same task and encouraged to learn from each other’s predictions. Concretely, the preceding network (i.e., the best-performing model from previous epochs) is used to extract spatio-temporal features from both query and support videos, and to compute an external similarity score with the current network. This stabilizes training, allows the present network to inherit the strengths of earlier models, and reduces overfitting. Experiments on benchmark datasets UCF-101, HMDB-51, and Kinetics show that our method significantly outperforms state-of-the-art few-shot video classification approaches. Ablation studies further confirm the contribution of the mutual learning mechanism. Our work advances the development of more efficient and accurate video classification systems for large-scale video data.