Accurate detection of cherry fruits is crucial for improving fruit quality, optimizing harvesting timing, and automating orchard management. However, due to factors such as small fruit size, diverse colors, complex backgrounds, and overlapping occlusions, detection still faced many challenges. To this end, this paper proposed a lightweight and improved model, ACS-YOLOv8n, based on the YOLOv8n architecture to enhance detection accuracy and efficiency. The model introduced an adaptive downsampling module (ADown) in the backbone network, which integrates the advantages of average pooling and max pooling, and combines multi-path convolution operations to effectively retain both global information and local salient features, thereby enhancing the detection capability of small targets. Meanwhile, the Context Guided Block (CGB) was used to replace the traditional C2f module, improving the model's ability to perceive the relationship between the target and the environment and enhancing the effect of multi-scale feature extraction through the joint extraction of local, contextual, and global features. In the neck part, the Slim-Neck structure based on GSConv and VoV-GSCSP was introduced to optimize the feature fusion method, significantly reduce computational redundancy, and balance speed and accuracy. The experimental results showed that ACS-YOLOv8n achieved 92.94% precision, 90.66% recall, and 96.57%mAP50 in complex orchard environments. Compared with YOLOv8n, the precision increased by 2.62 precentage point, the recall by 1.57 precentage point, and the average precision by 2.08 precentage point, while the number of model parameters and the computational volume were reduced by 31.90% and 33.33%, respectively, and the model size was 71.43% of the original model. The proposed model demonstrated good detection performance and robustness in complex environments, providing effective technical support for automatic cherry harvesting and intelligent orchard management.