At matched compute on an RTX 5070 Ti, the two families are even on accuracy: the mean gap across 7 pairs is 0.1 mAP points, split 3 to 3. RT-DETRv2 runs 11.7% faster on average in PyTorch and leads at the small end; RT-DETR leads at the large end. There is no single winner.
RT-DETR fields 7 measured variants; RT-DETRv2 fields 5. Both run the same COCO val2017 protocol at 640 px on an RTX 5070 Ti, and most variants share exact parameter counts, so the pairs line up cleanly.
| Model | Family | Params (M) | mAP@50-95 | FPS |
|---|---|---|---|---|
| RT-DETR-R18 | RT-DETR | 20.2 | 4979.0 | 47.2 |
| RT-DETRv2-R18 | RT-DETRv2 | 20.2 | 5075.0 | 51.1 |
| RT-DETR-R34 | RT-DETR | 31.4 | 5223.0 | 40.2 |
| RT-DETRv2-R34 | RT-DETRv2 | 31.4 | 5318.0 | 40.4 |
| RT-DETR-L | RT-DETR | 32.9 | 5577.0 | 27.7 |
| RT-DETR-R50m | RT-DETR | 36.6 | 5382.0 | 35.3 |
| RT-DETRv2-R50m | RT-DETRv2 | 36.6 | 5478.0 | 39.9 |
| RT-DETR-R50 | RT-DETR | 42.9 | 5588.0 | 29.8 |
| RT-DETRv2-R50 | RT-DETRv2 | 42.9 | 5571.0 | 32.1 |
| RT-DETR-X | RT-DETR | 67.4 | 5794.0 | 21.8 |
| RT-DETR-R101 | RT-DETR | 76.6 | 5677.0 | 23.6 |
| RT-DETRv2-R101 | RT-DETRv2 | 76.6 | 5677.0 | 25.7 |
Accuracy at matched compute
mAP is shown in percent. Across 7 nearest-parameter pairs the families split 3 and 3, with a mean gap of 0.1 mAP points. At matched R18 size RT-DETRv2 reaches 50.7 mAP against RT-DETR at 49.8; the order reverses in other pairs, which is why the aggregate is a tie.
Speed
RT-DETRv2 is faster at matched compute: 11.7% faster on average in PyTorch FP32. At matched R18 size it runs 51.1 FPS against RT-DETR at 47.2.
Where the frontier crosses
The accuracy lead depends on model size. RT-DETRv2 leads at the small-model end, RT-DETR leads at the large-model end. With the accuracy aggregate even, the speed edge and your target size decide the pick, not an overall ranking.
- 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 for small and mid-size deployments: it is faster at matched compute and leads on accuracy at the small end. Pick RT-DETR when you need a larger variant or a size RT-DETRv2 does not field, since it offers 7 variants to 5. 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.