YOLO26 vs. YOLOv8: A Comprehensive Architectural Benchmark of Next-Generation Real-Time Object Detection Models

This paper evaluates the performance of Ultralytics YOLO26 against YOLOv8, focusing on NMS-free architectures for real-time object detection in edge deployment scenarios. YOLO26 features a spectral-constrained CSP-Muon backbone and eliminates Distribution Focal Loss, achieving superior accuracy on the Pascal VOC dataset (0.635 mAP_50:95) while showing minimal performance differences on the VisDrone dataset. The study highlights that YOLOv8 maintains lower GPU inference latency (6.92 ms vs. 8.38 ms for YOLO26), indicating that NMS-free designs may not universally outperform traditional architectures, thus informing practitioners about architecture selection based on specific application needs.