Steel surface defect detection is essential for ensuring product quality and reducing costs in industrial manufacturing. Current methods have achieved promising performance, but challenges remain due to the diversity and complexity of defect types, sizes, and geometries. To address these challenges, this paper proposes MDF-YOLO, a lightweight and high-performance detection framework specifically designed for steel surface defect detection. To better handle complex defect geometries and scale variations, a Modulated Deformable Convolution Network (MDCNet) is integrated into the backbone to enable adaptive feature extraction. In addition, a Dynamic Coordination Mechanism (DyCM) is incorporated into the feature fusion neck between the backbone and detection head to strengthen multiscale feature fusion and spatial sensitivity. Furthermore, a Focaler-GIoU loss function is adopted to improve localization accuracy and reduce sample imbalance during bounding box regression. Experiments on the NEU-DET dataset demonstrate consistent performance gains. Compared with the baseline YOLOv8n, the proposed method achieves improvements of 3.9% in mAP@0.5, 3.2% in F1 score, 0.6% in recall, and 5.7% in precision. Overall, the proposed approach provides a robust, efficient, and deployment-friendly solution. It improves the performance of industrial surface defect detection systems.