Verdict

At the same backbone, RT-DETRv2 is the better pick: it scores up to 1.0 mAP point higher at R18, R34 and R50m, and it is as fast or faster at every shared size on an RTX 5070 Ti. Across all 7 nearest-parameter pairs the two families tie on accuracy, 3 wins each and a mean gap of 0.1 points, because RT-DETR also fields L and X variants that RT-DETRv2 does not. RT-DETRv2 runs 11.7% faster on average in PyTorch.

RT-DETR fields 7 measured variants; RT-DETRv2 fields 5. Five sizes share a backbone and an exact parameter count: R18, R34, R50m, R50 and R101. RT-DETR-L and RT-DETR-X have no RT-DETRv2 twin, so they pair with the nearest RT-DETRv2 model. All run the same COCO val2017 protocol at 640 px on an RTX 5070 Ti.

ModelFamilyParams (M)mAP@50-95FPS
RT-DETR-R18RT-DETR20.249.847.2
RT-DETRv2-R18RT-DETRv220.250.851.1
RT-DETR-R34RT-DETR31.452.240.2
RT-DETRv2-R34RT-DETRv231.453.240.4
RT-DETR-LRT-DETR32.955.827.7
RT-DETR-R50mRT-DETR36.653.835.3
RT-DETRv2-R50mRT-DETRv236.654.839.9
RT-DETR-R50RT-DETR42.955.929.8
RT-DETRv2-R50RT-DETRv242.955.732.1
RT-DETR-XRT-DETR67.457.921.8
RT-DETR-R101RT-DETR76.656.823.6
RT-DETRv2-R101RT-DETRv276.656.825.7
RT-DETR and RT-DETRv2 variant ladders interleaved by parameter count on NVIDIA RTX 5070 Ti, PyTorch FP32, batch 1. mAP in percent form. The family column shows where each frontier sits at every size.
Live chartverified data
Accuracy vs parameters on COCO val2017. Every RT-DETR and RT-DETRv2 variant highlighted against the full set.

Accuracy at matched compute

mAP is shown in percent. By backbone, RT-DETRv2 against RT-DETR: R18 50.7 vs 49.8, R34 53.2 vs 52.2, R50m 54.8 vs 53.8, R50 55.7 vs 55.9, and R101 56.8 for both. RT-DETRv2 wins three of the five shared sizes and gives back 0.2 points at R50. RT-DETR takes its other two wins with RT-DETR-L at 55.8 and RT-DETR-X at 57.9, which face the nearest RT-DETRv2 model. Across all 7 pairs the mean gap is 0.1 points.

Speed

RT-DETRv2 is as fast or faster at every shared backbone in PyTorch FP32: 51.1 vs 47.2 FPS at R18, 39.9 vs 35.3 at R50m, 32.1 vs 29.8 at R50 and 25.7 vs 23.6 at R101. At R34 the two are level, 40.4 vs 40.2. Averaged over all 7 pairs, RT-DETRv2 is 11.7% faster.

Where the frontier crosses

The accuracy lead follows model size. RT-DETRv2 leads at the small end, where its R18 and R34 models beat their RT-DETR twins. RT-DETR leads at the large end through RT-DETR-L and RT-DETR-X. With the aggregate even, your size class and the variants each family ships decide the pick.

Licensing
RT-DETR
Apache-2.0 (permissive)
RT-DETRv2
Apache-2.0 (permissive)
Commercial use
Both families ship under Apache-2.0

Which family to pick

Pick RT-DETRv2 if you would run R18, R34, R50m or R101: same parameter count, equal or higher mAP, and more FPS. Pick RT-DETR when you want RT-DETR-L or RT-DETR-X, sizes RT-DETRv2 does not field, or the 0.2 point edge at R50. Both ship under Apache-2.0. See the per-variant specs on the RT-DETR and RT-DETRv2 model pages.

Every number on this page comes from the verified dataset: same 500-image COCO val2017 slice, conf 0.001, IoU 0.6, max 300 detections, pycocotools mAP, identical protocol across all hardware and runtimes. The full protocol is on the methodology page. To rerun this comparison with your own filters, open compare. Accuracy is measured on LibreYOLO retrained checkpoints; other weight sources can yield different values.