Accuracy and CPU speed gains across lightweight YOLO generations under downstream uncertainty

Abstract Each new lightweight YOLO release reports higher COCO accuracy and lower CPU latency, and applied studies commonly build on the newest release. Recent multi-dataset comparisons show that a newer release need not win downstream, but they rarely separate a model difference from training-seed and test-set variability. We trained six nano-scale generations with three seeds on floating-debris and PASCAL VOC subsamples and three generations on VisDrone (49 runs), and compared them with paired test-image resampling. YOLO26n led YOLOv8n by 5.98 percentage points of [email protected]:0.95 on VOC but trailed it by 1.62 points on floating debris and 0.69 points on VisDrone, with all three bootstrap intervals excluding zero. The floating-debris deficit occurred in every object-size stratum and persisted under the default training recipe. These pairwise contrasts were stable, whereas the complete six-model ranking survived in only 55.1% and 16.4% of resamples. On three CPUs, randomized interleaved timing gave latency reductions of 1–12% from YOLOv8n to YOLO11n and 14–22% from YOLO11n to YOLO26n, against about 30% and 31% implied by the published tables. A release-level gain is therefore a hypothesis to test on the target data and processor. Pairwise intervals, not full rankings, should decide an upgrade or a baseline.