DEIM-L is the more accurate detector: 57.8 vs 55.7 mAP@50-95 against RT-DETRv2-R50 on an RTX 5070 Ti, a 2.1 point lead, with fewer parameters (31.24M vs 42.89M) and less VRAM (245 vs 327 MB). RT-DETRv2-R50 wins desktop-GPU speed, 32.1 vs 18.6 FPS. On the edge that reverses: DEIM-L is faster on both Jetson Orin and Raspberry Pi 5.
DEIM-L (31.24M parameters, Apache-2.0) and RT-DETRv2-R50 (42.89M, Apache-2.0) are both large detectors evaluated at 640 px on COCO val2017. DEIM-L is the smaller model. Both have verified rows on desktop GPU, Jetson Orin, and Raspberry Pi 5, so this comparison covers the deploy targets where the ranking changes.
| Metric | DEIM-L | RT-DETRv2-R50 |
|---|---|---|
| mAP@50-95 | 5783.0 | 5571.0 |
| mAP@50 | 7555.0 | 7284.0 |
| mAP small | 4515.0 | 3892.0 |
| FPS (mean) | 18.6 | 32.1 |
| Total ms/image | 53.76 | 31.15 |
| Inference ms | 46.27 | 26.73 |
| Peak VRAM (MB) | 245 | 327 |
| Params (M) | 31.2 | 42.9 |
| GFLOPs | 91.0 | 136.0 |
| Input size | 640 | 640 |
| License | Apache-2.0 | Apache-2.0 |
Accuracy
mAP is shown in percent form. DEIM-L measures 57.8 mAP@50-95 to RT-DETRv2-R50's 55.7, and the lead holds across runtimes. The gap is widest on small objects: 45.2 vs 38.9 mAP_small, a 16.02% relative gap. DEIM-L is also the more compute-efficient model, extracting 54.88% more mAP per GFLOP.
Speed
On the RTX 5070 Ti in PyTorch, RT-DETRv2-R50 runs 32.1 FPS to DEIM-L's 18.6, so it is 42.07% faster. The margin holds under conversion: 66.0 vs 50.5 FPS on ONNX Runtime and 88.5 vs 75.7 FPS on TensorRT FP32.
The speed verdict does not travel to the edge. On Jetson Orin DEIM-L runs 4.8 FPS to RT-DETRv2-R50's 4.3, and on Raspberry Pi 5 it leads 0.4 to 0.3 FPS. If your target is a Jetson or a Pi, the desktop numbers point the wrong way.
- DEIM-L license
- Apache-2.0
- RT-DETRv2-R50 license
- Apache-2.0
- DEIM-L release
- 2024-12-05
- RT-DETRv2-R50 release
- 2024-05-01
- Evaluated weights
- LibreYOLO retrained checkpoints
When to pick which
Pick DEIM-L for accuracy, memory, compute efficiency, or edge serving, where it leads at every object size and is faster on both Jetson Orin and Raspberry Pi 5. Pick RT-DETRv2-R50 for desktop or server GPU throughput, where it is faster on every GPU runtime measured. Both are Apache-2.0, so licensing does not force the choice.
Try both with LibreYOLO
# pip install libreyolo
from libreyolo import LibreYOLO
deim = LibreYOLO("LibreDEIMl.pt") # 31.24M params, 57.8 mAP@50-95
rtdetr = LibreYOLO("LibreRTDETRv2r50.pt") # 42.89M params, 55.7 mAP@50-95
results = deim("image.jpg")
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.