Sheep face recognition technology is critical in key areas such as individual sheep identification and behavior monitoring. Existing sheep face recognition models typically require high computational resources. When these models are deployed on mobile or embedded devices, problems such as reduced model recognition accuracy and increased recognition time arise. To address these problems, an improved Parameter Fusion Lightweight You Only Look Once (PFL-YOLO) sheep face recognition model based on YOLOv8n is proposed. In this study, the Efficient Hybrid Conv (EHConv) module is first integrated to enhance the extraction capability of the model for sheep face features. At the same time, the Residual C2f (RC2f) module is introduced to facilitate the effective fusion of multi-scale feature information and improve the information processing capability of the model; furthermore, the Efficient Spatial Pyramid Pooling Fast (ESPPF) module was used to fuse features of different scales. Finally, parameter fusion optimization work was carried out for the detection head, and the construction of the Parameter Fusion Detection (PFDetect) module was achieved, which significantly reduced the number of model parameters and computational complexity. The experimental results show that the PFL-YOLO model exhibits an excellent performance-efficiency balance in sheep face recognition tasks: mAP@50 and mAP@50:95 reach 99.5% and 87.4%, respectively, and the accuracy is close to or equal to the mainstream benchmark model. At the same time, the number of parameters is only 1.01 M, which is reduced by 45.1%, 83.7%, 66.6%, 71.4%, and 61.2% compared to YOLOv5n, YOLOv7-tiny, YOLOv8n, YOLOv9-t, and YOLO11n, respectively. The size of the model was compressed to 2.1 MB, which was reduced by 44.7%, 82.5%, 65%, 72%, and 59.6%, respectively, compared to similar lightweight models. The experimental results confirm that the PFL-YOLO model maintains high accuracy recognition performance while being lightweight and can provide a new solution for sheep face recognition models on resource-constrained devices.