A new academic paper traces five years of Ultralytics' YOLO object detectors, ending with an unreleased model called YOLO27.
The review, posted to the arXiv preprint server, walks through the family from YOLOv5's modular PyTorch design through YOLOv8's anchor-free detector, YOLO11's efficiency-focused updates, and YOLO26's changes, including dropping distribution focal loss and adding NMS-free inference. It caps the list with YOLO27, described as a two-tier design: compact YOLO27n and YOLO27s models use streamlined CNNs, while the larger YOLO27m and YOLO27l use transformer-based decoding to skip traditional NMS post-processing entirely. The paper's cited COCO benchmarks put accuracy between 42.3 and 61.2 mAP depending on model size and input resolution, with the smallest models claiming latency under a millisecond on TensorRT 11.
YOLO has been the go-to fast object-detection architecture since 2020, and Ultralytics has kept shipping new versions faster than most teams can properly vet them. A paper that lines up five generations side by side makes that pace visible, and since YOLO27 doesn't appear to be publicly released yet, this reads as an early look rather than a launch.
Every number here comes from the paper itself, not from outside testing, so treat YOLO27's claimed 61.2 mAP as a preview until someone else reproduces it.