This study presents a lightweight SOD-YOLOv5n model for real-time winter jujube detection and counting, deployed on Android devices. The SOD-YOLOv5n model is an improvement over the YOLOv5n model, incorporating SPD-Conv layers to enhance small object detection and low-resolution image handling. The CARAFE module is used for upsampling, and GSConv modules replace some conv modules in the neck to reduce model complexity and improve accuracy. The model is quantized using float16 to reduce its size to 3.64 MB, making it suitable for deployment on Android devices. An app named JujubeDetector is developed to perform real-time detection and counting, with a detection time of 30-90 ms. Experiments in winter jujube orchards show that the SOD-YOLOv5n model achieves high accuracy, with improvements of 2.40%, 1.80%, and 3.00% in precision, recall, and mAP, respectively, compared to the YOLOv5n model. The model also reduces RMSE and MAPE by 9.11% and 5.30%, respectively. The JujubeDetector app demonstrates effective performance in complex environments, with a root mean square error (RMSE) of 1.46, coefficient of determination (R²) of 0.97, and detection time of 30-90 ms. This approach provides a practical solution for real-time winter jujube detection and counting, supporting yield estimation and offering a reference for detecting other small target fruits.This study presents a lightweight SOD-YOLOv5n model for real-time winter jujube detection and counting, deployed on Android devices. The SOD-YOLOv5n model is an improvement over the YOLOv5n model, incorporating SPD-Conv layers to enhance small object detection and low-resolution image handling. The CARAFE module is used for upsampling, and GSConv modules replace some conv modules in the neck to reduce model complexity and improve accuracy. The model is quantized using float16 to reduce its size to 3.64 MB, making it suitable for deployment on Android devices. An app named JujubeDetector is developed to perform real-time detection and counting, with a detection time of 30-90 ms. Experiments in winter jujube orchards show that the SOD-YOLOv5n model achieves high accuracy, with improvements of 2.40%, 1.80%, and 3.00% in precision, recall, and mAP, respectively, compared to the YOLOv5n model. The model also reduces RMSE and MAPE by 9.11% and 5.30%, respectively. The JujubeDetector app demonstrates effective performance in complex environments, with a root mean square error (RMSE) of 1.46, coefficient of determination (R²) of 0.97, and detection time of 30-90 ms. This approach provides a practical solution for real-time winter jujube detection and counting, supporting yield estimation and offering a reference for detecting other small target fruits.