LibreYOLO Benchmarks

Compare COCO accuracy, generalization and inference speed.

COCO

Accuracy against size

Compare COCO accuracy against parameter count, with YOLOv9 and RF-DETR highlighted. Data from Vision Analysis.

Open the full leaderboard

RF100-VL

Generalization beyond COCO

RF100-VL measures how object detectors adapt to new domains through fine-tuning on 100 real-world datasets.

17 configurations · mean AP50:95
Mean AP50:95485256606401020304050Total parameters (M)NSMLTSMNanoTinySMSMLSMLoRA
Hover or tap a point for model details.
RF100-VL benchmark reportPublic training and evaluation harness
Methodology and caveats

Each configuration trains separately on each dataset. Checkpoints are selected on validation and scored on test with pycocotools at maxDets 500. The result is the unweighted mean AP50:95 across all 100 datasets.

Lines connect model sizes. Recipes and resolutions vary; LoRA is shown separately.

These are single-seed results with family-specific recipes, resolutions and precision. The report documents the methodology, model comparisons and implementation limits.

* YOLOv9 scores predate training fixes. YOLO-NAS uses separately licensed, non-commercial pretrained weights.

Inference speed on your hardware

Compare latency by device, runtime and precision.

See latency on every board