The rapid expansion of photovoltaic (PV) systems underscores the need for efficient defect classification frameworks to ensure their reliable and uninterrupted operation. Although Vision Transformer (ViT) have demonstrated exceptional performance in image classification tasks, their high computational complexity makes them unsuitable for real-time deployment on resource-constrained edge devices commonly found in PV systems. To address this limitation, we propose a lightweight knowledge distillation (KD) framework tailored for multi-class solar cell defect classification. In this approach, a compact MobileNetV3 student model learns from a powerful ViT-B-8 teacher, effectively transferring knowledge to achieve high classification performance with minimal resource requirements. The distilled MobileNetV3 achieves a test accuracy of 97.9% and a validation accuracy of 97.54%, closely matching the teacher’s performance. Moreover, it reduces the number of parameters by 93.62%, inference time by 79.01%, and GFLOPs by 92.16%, highlighting its efficiency. The framework’s effectiveness was validated through metrics including accuracy, precision, recall, F1-score, and inference speed. These results demonstrate that the proposed KD framework offers an excellent trade-off between accuracy and computational efficiency, making it highly suitable for real-time defect monitoring in PV systems deployed on edge devices.