Cervical cancer is a major global health issue, especially in low-resource settings where Pap smear diagnosis is limited by manual interpretation and a lack of automated tools. While deep learning shows potential for classifying cytology images, current models struggle with high computational needs, poor performance on class-imbalanced datasets, and limited interpretability, which impede their use in clinical settings. To address these problems, we present CerviLightStack, a lightweight stacking ensemble framework for reliable cervical cancer classification using Pap smear images. This model combines four efficient deep learning backbones - MobileOne-S, EfficientViT-M0, EdgeNeXt-Small, and MobileViT v2XXS - with a Random Forest meta-learner to effectively capture diverse cytological features. We evaluated it using two benchmark datasets: SIPaKMeD, with 4,049 images across 5 classes, and Herlev, featuring 917 images across 7 classes, both restructured into clinically relevant binary and multiclass categories. We used comprehensive data preprocessing and advanced augmentation methods like Mixup and CutMix to address data imbalance. Our evaluation metrics included accuracy, specificity, PR-AUC, and MCC. CerviLightStack achieved 99.41 % accuracy on the SIPaKMeD dataset and 99.76 % on the Herlev dataset, surpassing current state-of-theart models. We developed an explainable web application using Grad-CAM for real-time inference and transparent decisionmaking. Our system shows high accuracy, efficiency, and clinical utility, making it a scalable solution for cervical cancer screening in real-world settings.