In order to effectively lightweight the leaf disease detection model under the premise of ensuring the recognition performance, a model lightweight method was constructed based on trunk replacement, model pruning and knowledge distillation technology, and a lightweight test was carried out on the leaf yellow leaf curl disease detection model based on YOLO v5s. Firstly, the main body of the model was reduced by replacing the YOLO v5s trunk with the common lightweight convolutional neural networks (LCNN) with excellent performance. Then, the unimportant channels were screened and deleted by using the sparse training of the model and the distribution of the scaling factors in the batch normalization layer. Finally, by fine-tuning retraining and knowledge distillation, the model accuracy was adjusted to a level close to that before pruning. The experimental results showed that the accuracy, recall and mean average accuracy of the lightweight model were 91.3%, 87.4% and 92.7%, respectively. The memory consumption of the model was 1.4MB, and the detection frame rate of the desktop was 81.0f/s. The detection frame rate of the mobile terminal was 1.2f/s. Compared with the original YOLO v5s leaf disease detection model, the accuracy, recall and average accuracy were reduced by 3.7 percentage points, 4.6 percentage points and 2.7 percentage points, and the memory consumption was only 10% of that before processing. The frame rate of the desktop and mobile terminal detection was increased by nearly 27% and 33%, respectively. The proposed method can effectively reduce the weight of the model under the premise of keeping the performance, which provided a theoretical basis for the deployment of mobile leaf disease detection.