To address the challenges posed by the diverse growth states of Rosa davurica Pall in complex environments, significant scale variation of targets, and the stringent requirements for lightweight and real-time detection in resource-constrained scenarios, this work proposes CMS-YOLO, a lightweight functionality-enhanced detection model based on YOLOv8n. The model first integrates a novel CSPHet module-constructed by combining dual-branch design with a heterogeneous convolutional mechanism-to enhance feature extraction capacity while minimizing redundant computation. Subsequently, a Cross-Channel Feature Fusion Module (CCFM) is introduced into the neck network, effectively reducing the parameter count while preserving high-efficiency information flow. Furthermore, the SENetV2 attention mechanism is embedded across multi-scale paths (P3/P4/P5) to bolster the model’s representational strength and robustness for multi-scale features. Empirical results demonstrate that CMS-YOLO achieves a mAP@0.5 of 93% and a precision of 90.4%, with an inference speed of 182 FPS. The model’s parameter size and storage footprint are constrained to merely 1.60M and 3.36 MB, respectively. These results underscore CMS-YOLO’s superiority over state-of-the-art counterparts in striking a favorable balance among accuracy, efficiency, and complexity, while also confirming its successful deployment on desktop platforms, showcasing strong practicality and potential for widespread application.