Automated detection of surface defects in steel is critical for ensuring product quality and operational efficiency in modern manufacturing. This study presents a unified framework for real-time steel surface inspection using five lightweight YOLO (You Only Look Once) nano architectures: YOLOv5n, YOLOv8n, YOLOv11n, YOLOv12n, and YOLOv13n. Unlike prior studies that focused on optimizing individual models, this work conducts a cross-generation comparative analysis, introducing the first evaluation of YOLOv13 for steel surface defect detection. The proposed framework integrates systematic dataset preparation, model training on benchmark and custom industrial datasets, and detailed performance assessment under stringent IoU thresholds and real-time inference conditions. Experimental results reveal that inference performance depends more on architectural efficiency, hardware utilization, and software-level optimization than on model size alone. Basic augmentation techniques, such as flipping and rotation, further enhance the detection of small and hard-to-capture defects. Among all models, YOLOv13n achieves the fastest inference of 303.03 FPS with competitive accuracy, demonstrating exceptional suitability for real-time, edge-based industrial deployments. The findings provide valuable empirical insights for selecting efficient architectures in automated visual inspection systems.