To improve the efficiency and accuracy of cherry tomato detection for harvesting robots in greenhouses, we propose a lightweight detection network based on YOLOv8 with enhanced feature fusion and loss function (termed LEFF-YOLO). Firstly, the SiMAM parameter-free attention mechanism is introduced, and a SiMAM-C2f module is incorporated into the neck network to enhance the model’s feature fusion capability, enabling it to capture key features of cherry tomatoes more accurately; Then, the VanillaNet module is utilzied to reconstruct the backbone network, reducing the number of model parameters and improving detection speed; Finally, a new bounding box loss function was designed, which enhances the model’s stability and detection accuracy. Experimental results show that compared to YOLOv8, the average detection accuracies (mAP50 and mAP50-95) of LEFF-YOLO increase by 2.7% and 2.9% respectively, its number of parameters decreased by 32%, and its computational cost reduces by 10.5%. The improved lightweight network can meet the requirements of real-time and efficient cherry tomato detection for harvesting robots.