Printed circuit board (PCB) defect detection is crucial for ensuring product quality and reliability. YOLOv10 represents the latest state-of-the-art in object detection models, distinguishing itself by eliminating post-processing to enhance real-time processing speed compared to all its predecessors, making it ideal for resource-constrained manufacturing environments. However, YOLOv10 shows limitations in detecting tiny defects. To address this, we introduce Dynamic YOLO with Multi-scale attention (DM-YOLO), a novel model built upon improved YOLOv10 that integrates dynamic convolution in the backbone, dynamic upsampling in the neck, and a dynamic loss function to enhance performance at both feature extraction and fusion level. Additionally, DM-YOLO introduces a novel Partial Efficient Multi-scale Attention (PEMA) mechanism to replace traditional multi-head self-attention, further enhancing feature extraction. Beyond detection, precise defect localization is crucial for potential automated repair processes, emphasizing the importance of mAP@50–95. DM-YOLO achieves 98.5% and 58.7% on TDD dataset, and 97.7% and 76.8% on DeepPCB for mAP@50 and mAP@50–95 metrics, respectively. These results reflect improvements of up to 4.3% in mAP@50–95 and up to 1.5% in mAP@50 compared to YOLOv10. Moreover, it reduces computational complexity from 8.2 to 7.7 GFLOPS, enhancing both precision and real-time performance in PCB defect detection and localization.