Continuous blood pressure (BP) monitoring holds potential in preventing and detecting cardiovascular disease (CVD). Photoplethysmography (PPG)-based BP measuring systems with noninvasive and continuous properties are of enormous research value in biomedical science. However, mainstream technologies such as Transformer are not viable for implementation in compact devices due to their significant processing complexity. Additionally, the unique individual variability of biosignals limits the generalization performance of the model. In this work, we propose a personalized modeling approach employing a lightweight convolutional neural network(CNN) that streamlines the structure while preserving the model’s ability for long-term predictions. We leverage continuous records from 413 and 30 subjects extracted from MIMIC-III for model pretraining and personalization. The final results demonstrate that, calibrated with only 30 labeled windows for the target subject, the personalized model achieves an estimation error of -0.218±7.657 mmHg for systolic BP (SBP) and -0.261±4.898 mmHg for diastolic BP (DBP). This performance meets the standard requirements of the Association for the Advancement of Medical Instrumentation (AAMI). The lightweight structure and personalized modeling approach improve the accuracy of BP estimation while reducing the difficulty of deployment.