Best YOLOv8 Alternatives in 2026
아티클은 영어로 작성되어 있습니다.
COCO is the usual way to pick a detector. Another angle is how well the model fine-tunes on other data. RF100-VL measures that: a COCO-pretrained checkpoint is trained for 100 epochs on each of 100 datasets, then scored on each test split. The number is the mean mAP50-95.
YOLOv8m is 56.9 there, from the paper appendix. Two models we ran score higher: RF-DETR and YOLO-NAS. Raw runs are in huggingface.co/datasets/LibreYOLO/rf100-vl-results.
RF-DETR
RF-DETR-S is 60.41 and RF-DETR-M is 61.13 in LibreYOLO's campaign. Roboflow reports 60.2 and 61.2. Nano through Large are Apache-2.0.
from libreyolo import LibreYOLO
model = LibreYOLO("LibreRFDETRs.pt")
model.train(data="my-dataset.yaml", epochs=100, imgsz=512, batch=8)
Docs | Weights | Run 20260814-rfdetr-s-1b1190ee
YOLO-NAS
YOLO-NAS-S is 58.00, the convolutional option if you want a YOLO-style CNN. Deci's published weights are non-commercial. LibreYOLO hosts none of them. Details: YOLO-NAS is still maintained.
from libreyolo import LibreYOLO
model = LibreYOLO("LibreYOLONASs.pt")
model.train(data="my-dataset.yaml", epochs=100, imgsz=640, batch=16)
Docs | Run 20260809-yolonas-s-02926964
Install with pip install "libreyolo[rfdetr]" for RF-DETR, or pip install libreyolo for YOLO-NAS. Campaign page: RF100-VL.