The answer depends on your runtime. In PyTorch on an RTX 5070 Ti, RF-DETR-L is 0.8 mAP points more accurate, 58.6 vs 57.8 mAP@50-95, and 25.84% faster, 25.1 vs 18.6 FPS. Convert the models and DEIM-L takes the speed lead: it is faster on ONNX Runtime, TensorRT FP32, and Jetson Orin, and it leads on small objects. The resolutions differ: DEIM-L runs at 640 px and RF-DETR-L at 704 px, each at its author default.
DEIM-L (31.24M parameters) and RF-DETR-L (33.93M) both ship under Apache-2.0 and evaluate on COCO val2017 at different input sizes: DEIM-L at 640 px, RF-DETR-L at 704 px, each at the author default. Both have verified rows on desktop GPU, Jetson Orin, and Raspberry Pi 5.
| Metric | DEIM-L | RF-DETR-L |
|---|---|---|
| mAP@50-95 | 57.8 | 58.6 |
| mAP@50 | 75.5 | 76.0 |
| mAP small | 45.2 | 40.9 |
| FPS (mean) | 18.6 | 25.1 |
| Total ms/image | 53.76 | 39.87 |
| Inference ms | 46.27 | 21.95 |
| Peak VRAM (MB) | 245 | 223 |
| Params (M) | 31.2 | 33.9 |
| GFLOPs | 91.0 | 340.0 |
| Input size | 640 | 704 |
| License | Apache-2.0 | Apache-2.0 |
Accuracy
mAP is shown in percent form. RF-DETR-L measures 58.6 mAP@50-95 to DEIM-L's 57.8, a 0.8 point lead. DEIM-L takes small objects, 45.2 vs 40.9 mAP_small, a 10.49% relative gap.
Speed
In PyTorch on the RTX 5070 Ti, RF-DETR-L runs 25.1 FPS to DEIM-L's 18.6. Convert either model and DEIM-L takes over: 75.7 vs 45.1 FPS on TensorRT FP32, 50.5 vs 43.8 on ONNX Runtime. RF-DETR-L reaches its accuracy on 340 GFLOPS, so DEIM-L extracts 269.19% more mAP per GFLOP.
The PyTorch speed win is narrow ground. DEIM-L is faster on ONNX Runtime, TensorRT FP32, and Jetson Orin, where it runs 4.8 FPS to RF-DETR-L's 4.0.
- DEIM-L license
- Apache-2.0
- RF-DETR-L license
- Apache-2.0
- DEIM-L release
- 2024-12-05
- RF-DETR-L release
- 2026-02-23
- Evaluated weights
- LibreYOLO retrained checkpoints
When to pick which
Pick RF-DETR-L for PyTorch deployment and for the top mAP at 704 px. Pick DEIM-L for converted runtimes and Jetson, where it is faster, stronger on small objects, and far more compute-efficient. Both ship under 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
rfdetr = LibreYOLO("LibreRFDETRl.pt") # 33.93M params, 58.6 mAP@50-95
results = rfdetr("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.
