On an RTX 5070 Ti in PyTorch, DEIM averages 4.3 mAP points higher than RT-DETR at matched parameter counts, winning all 3 nearest-params pairings. RT-DETR is 41.6% faster on average. DEIM leads accuracy across the range; RT-DETR wins on throughput.
DEIM fields 5 measured variants and RT-DETR fields 7, all on the same COCO val2017 protocol on an RTX 5070 Ti. Three pair closely by parameter count, giving 3 matched comparisons.
| Model | Family | Params (M) | mAP@50-95 | FPS |
|---|---|---|---|---|
| DEIM-N | DEIM | 3.8 | 4679.0 | 37.0 |
| DEIM-S | DEIM | 10.3 | 5210.0 | 33.4 |
| DEIM-M | DEIM | 19.6 | 5549.0 | 28.4 |
| RT-DETR-R18 | RT-DETR | 20.2 | 4979.0 | 47.2 |
| DEIM-L | DEIM | 31.2 | 5783.0 | 18.6 |
| 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 |
| DEIM-X | DEIM | 62.6 | 5961.0 | 15.0 |
| 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 DEIM variant is paired with the nearest RT-DETR variant by parameter count. Across the 3 pairings DEIM averages 4.3 mAP points higher and wins all three. DEIM-X reaches 59.6 mAP at the top; RT-DETR-R101 reaches 56.8. At the small end DEIM-N is a 3.78M model at 46.8 mAP, against RT-DETR-R18 at 20.18M and 49.8 mAP.
Speed
Averaged across the matched pairs, RT-DETR is 41.6% faster than DEIM in PyTorch. The gap is widest at the DEIM-L pairing and narrowest at the flagship pairing.
- DEIM license
- Apache-2.0
- RT-DETR license
- Apache-2.0
- Both families
- permissive, cleared for commercial use
Which family to pick
Pick DEIM for accuracy per parameter: it wins every matched pairing and reaches 59.6 mAP. Pick RT-DETR for throughput, at 41.6% faster on average, and for the wider ladder of 7 variants. Both ship under Apache-2.0, so licensing does not force the call.
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.