Cervical cancer, primarily caused by human papillomavirus (HPV) infection, is the second most common cancer in women globally. Early detection via cytological screening is essential to reduce mortality. However, manual screening methods are time-consuming and error-prone, underscoring the need for automated, efficient diagnostic tools. This study presents a novel lightweight deep learning architecture for cervical cancer classification in cytological images, designed to overcome the computational demands of standard Vision Transformer (ViT) and Swin Transformer models while preserving high diagnostic accuracy. We propose a hybrid MobileNetV3-Small + MobileViT model that combines the local feature extraction strengths of convolutional neural networks with the global contextual modeling of vision transformers. The model integrates Efficient Channel Attention (ECA), pyramidal feature fusion, and ShakeDrop regularization to enhance performance. Evaluated on the SipakMed dataset, the model achieved approximately 88% accuracy with only 1.17M parameters, 98% reduction in model size compared to ViT and Swin Transformers (86M approximately). Training time was also significantly reduced (around 15 s per epoch), enabling fast and efficient deployment. This lightweight hybrid model offers a practical solution for cervical cancer screening in resource-limited settings, supporting mobile health applications and point-of-care diagnostics with minimal compromise in accuracy.