Purpose/Significance To explore lightweight model suitable for blood cell classification, so as to facilitate rapid auxiliary diagnosis of blood diseases. Method/Process A lightweight model YOLOv8-SCG is designed based on the YOLOv8 architecture. Comparative experiments and ablation experiments are conducted on the public blood cell count and detection dataset (BCCD), and generalization testing is completed based on the Leukemia-cell dataset. Result/Conclusion The detection accuracy and lightweight performance of the YOLOv8-SCG model are significantly improved compared with mainstream models such as YOLOv8n and YOLOv11n. The precision, recall and mAP on the Leukemia-cell pathological abnormal blood cell dataset reach 94.3%, 95.6% and 95.4% respectively, with excellent generalization ability, providing an efficient and feasible method for the automatic classification of blood cells.