At matched compute on an RTX 5070 Ti, YOLOv9 leads RT-DETR on both axes. It averages 4.0 mAP points higher across 4 nearest-parameter pairs and wins all 4, and it runs 7.2% faster on average in PyTorch. RT-DETR's advantage here is range, not results: it fields more variants, but none of its matched pairs beats YOLOv9.
RT-DETR fields 7 measured variants; YOLOv9 fields 4. Both run the same COCO val2017 protocol at 640 px on an RTX 5070 Ti, so the two ladders align by parameter count.
| Model | Family | Params (M) | mAP@50-95 | FPS |
|---|---|---|---|---|
| YOLOv9-T | YOLOv9 | 2.0 | 4178.0 | 31.9 |
| YOLOv9-S | YOLOv9 | 7.2 | 5045.0 | 30.8 |
| YOLOv9-M | YOLOv9 | 20.1 | 5610.0 | 37.1 |
| RT-DETR-R18 | RT-DETR | 20.2 | 4979.0 | 47.2 |
| YOLOv9-C | YOLOv9 | 25.5 | 5709.0 | 42.3 |
| RT-DETR-R34 | RT-DETR | 31.4 | 5223.0 | 40.2 |
| RT-DETR-L | RT-DETR | 32.9 | 5577.0 | 27.7 |
| RT-DETR-R50m | RT-DETR | 36.6 | 5382.0 | 35.3 |
| RT-DETR-R50 | RT-DETR | 42.9 | 5588.0 | 29.8 |
| RT-DETR-X | RT-DETR | 67.4 | 5794.0 | 21.8 |
| RT-DETR-R101 | RT-DETR | 76.6 | 5677.0 | 23.6 |
Accuracy at matched compute
mAP is shown in percent. Each RT-DETR model is paired with the nearest YOLOv9 model by parameter count. YOLOv9 wins all 4 pairs and averages 4.0 mAP points higher. The pattern holds through the mid range: YOLOv9-M reaches 56.1 mAP against RT-DETR-R18 at 49.8 in the same parameter class.
Speed
YOLOv9 is also faster at matched compute: 7.2% faster on average in PyTorch FP32. RT-DETR's throughput lead shows up only against models in a different size class, not at matched parameters.
- RT-DETR
- Apache-2.0 (permissive)
- YOLOv9
- MIT (permissive)
- Commercial use
- Both families ship under permissive licenses
Which family to pick
Pick YOLOv9 across the measured range: it is more accurate at every matched pair and faster on average. Pick RT-DETR only when you need a size YOLOv9 does not field, since it offers 7 variants to YOLOv9's 4. Both ship under permissive licenses, so licensing does not force the call. See the per-variant specs on the RT-DETR and YOLOv9 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.