Using real-time and precise detection methods for maize leaf disease can significantly reduce economic losses in agriculture. Practical implementation often faces challenges such as the large volume of leaf disease data, low identification accuracy, and inefficiencies in production environments. To address these issues, this study introduces YOLO-MSM, a maize leaf disease detection algorithm that integrates multi-scale variable kernel convolution. In the YOLO MSM algorithm, we introduce an innovative convolutional method, MKConv (Multi-scale Variable Kernel Convolution), which offers diverse parameter configuration options and adapts flexibly to sample shapes with specific data characteristics. This design significantly enhances the network's overall performance. Additionally, to highlight critical features and mitigate the influence of environmental noise, we develop the C2f-SK module, leveraging the SK (Selective Kernel) attention mechanism to optimize feature extraction and representation. The loss function is optimized using MPDIoU (Minimum Point Distance Intersection over Union) to enhance the algorithm's capability in accurately locating densely occluded targets. The findings from the experiments indicate that the YOLO MSM algorithm reaches a real-time detection rate of 279.56 fps. In comparison to the baseline algorithm, the algorithm improves the precision and recall by 0.66% and 1.61%, respectively. Moreover, YOLO MSM algorithm is effectively lightweight compared to the series of cutting-edge algorithm models, which are only 5.4 MB in size, and the number of parameters and Flops are also reduced significantly. Therefore, YOLO MSM algorithm has an obvious light-weight advantage, which can achieve a good balance between precision and speed, and lay a theoretical foundation for identifying and detecting leaf disease on mobile devices.