With the rapid development of the railway industry, ensuring the operational safety of railways places increasingly stringent demands on the efficiency and accuracy of rail surface defect detection. To address the low efficiency of traditional inspection methods and the imbalance between computational complexity, model lightweighting, and detection accuracy in existing deep learning–based approaches, this paper proposes a lightweight rail surface defect detection model termed RSD-YOLO. Built upon YOLO11n, the proposed model introduces the HSFPN module into the neck to enhance multi-scale feature fusion while maintaining a lightweight architecture. In the backbone network, ADown downsampling and the RFE receptive field enhancement module are integrated to improve the detection performance of small-scale rail defects. Furthermore, an Ultra-LSPPF lightweight pooling module and a redesigned detection head based on MBConv convolutions are employed to systematically reduce model complexity while enhancing overall performance. Experimental results on the rail surface defect dataset demonstrate that, compared with the original YOLO11n, RSD-YOLO reduces the number of parameters, computational cost, and model size by 62%, 57%, and 58%, respectively, while achieving a 4.8% improvement in detection accuracy. The proposed model attains an mAP@0.5 of 79.2% and an inference speed of 142 FPS. These results indicate that RSD-YOLO effectively balances detection accuracy and inference efficiency while achieving significant lightweighting, making it well suited for real-time intelligent rail surface defect detection with substantial engineering application value.