Accurate and timely detection of lychee leaf diseases is essential for improving crop yield and minimizing economic losses. Existing approaches, which are largely based on standalone convolutional neural networks (CNNs), often fail to effectively capture both local visual patterns and global contextual relationships, limiting their classification performance and real-world applicability. To overcome these limitations, this study proposes LMViT, a lightweight hybrid framework that combines MobileNet for efficient local feature extraction with a Tiny Vision Transformer (ViT) to model long-range contextual dependencies, enabling accurate and computationally efficient disease classification. A high-quality dataset consisting of 3,766 high-resolution lychee leaf images across six disease categories was collected from Biral, Dinajpur, Bangladesh. The performance of four state-of-the-art CNN models‐‐EfficientNetV2S, ResNet50, Xception, and MobileNet‐‐was evaluated, where MobileNet achieved the best individual accuracy of 99.33 %. The proposed LMViT model further improved performance, achieving an overall accuracy of 99.88 %. To ensure robustness and generalization, a 5-fold stratified cross-validation was conducted, resulting in a mean validation accuracy of 99.87%(±0.13%), with corresponding mean precision, recall, and F 1-score of 9 9. 8 9 %,99.87%, and 99.88 %, respectively. Additionally, LMViT was deployed in a web-based application to enable real-time disease detection and provide actionable treatment recommendations. Experimental results demonstrate that LMViT offers a reliable, efficient, and practical solution for automated lychee leaf disease detection across all six disease classes.