Skip to content
ModelsThatRun

Object detection: Real-time, closed set

A fixed label set such as the 80 COCO classes, fast enough for video.

1251020501003035404550556065Latency on RTX 4090, ms, log scale (measured)box AP, COCO val2017 (measured)nsmlxsmlxsmlxnanosmallmediumlargesmlxnsmlxnsmlxnsmlx
● Commercial OK● Conditional● Non-commercialBest trade-off so farInput sizes differ: compare latency with care

All real-time, closed set models

Each model runs at its official default input, shown under its latency. A smaller input is less work, so compare latencies across input sizes with care.

ModelAccuracy (paper under it)4090 latency (input under it)VRAM peakLicense
D-FINE5 sizes: n, s, m, l, x42.6 box AP42.8 COCO AP val paper21.9 ms640x640563 MiBCommercial OK
DEIMv24 sizes: s, m, l, x50.6 box AP50.9 COCO AP paper27.4 ms640x640575 MiBNon-commercial
GTR4 sizes: s, m, l, x53.3 box AP53.6 COCO AP paper34.0 ms640x640636 MiBCommercial OK
RF-DETR4 sizes: nano, small, medium, large47.9 box AP48.4 COCO AP50:95 paper9.8 ms384x384803 MiBCommercial OK
RT-DETRv44 sizes: s, m, l, x49.6 box AP49.8 COCO AP paper24.5 ms640x640639 MiBCommercial OK
YOLO115 sizes: n, s, m, l, x39.1 box AP39.5 COCO mAP50-95 paper14.0 ms640x448572 MiBConditional
YOLO265 sizes: n, s, m, l, x40.5 box AP40.9 COCO mAP50-95 paper14.2 ms640x448571 MiBConditional
YOLOv85 sizes: n, s, m, l, x37.0 box AP37.3 COCO mAP50-95 paper11.6 ms640x448514 MiBConditional