Verdict

On the Raspberry Pi 5, permissive models match the restricted ones on accuracy and, from the small class up, run at least as fast. RF-DETR-M (Apache-2.0) reaches 57.4 mAP at 0.8 FPS against YOLO26-X (AGPL-3.0) at 56.5 mAP and 0.4 FPS. RF-DETR-N lands within 0.9 points of YOLO11-L, YOLOv8-L and YOLO-NAS-S at 2.1 FPS, against 0.9 FPS or less. Only at the nano end do YOLO26-N and YOLO11-N keep a speed lead.

License terms decide whether a model can ship in a closed commercial product. YOLO-NAS carries a non-permissive license from its authors. The Ultralytics models (YOLOv5u, YOLOv8, YOLO11, YOLO26) ship under AGPL-3.0, a copyleft license with source-release obligations. MIT and Apache-2.0 carry no such obligation. This guide covers 22 restricted models and 48 permissive ones on the same Raspberry Pi 5 protocol in PyTorch FP32. mAP is shown in percent form throughout.

This is not legal advice. License names come from each model's upstream repository. Confirm the exact terms with the upstream project and your own counsel before you ship.

Large: RF-DETR-M and D-FINE-S

RF-DETR-M is 0.9 points ahead of YOLO26-X and runs at twice its frame rate, 0.8 vs 0.4 FPS. RF-DETR-S reaches 55.1 mAP at 1.0 FPS against YOLO-NAS-M at 55.4 and 0.5 FPS. D-FINE-S reaches 53.4 mAP at 1.1 FPS. YOLO26-L sits 0.4 points higher at 0.7 FPS, and YOLO11-X reaches 53.6 at 0.4 FPS.

MetricYOLO26-XYOLO-NAS-MRF-DETR-MRF-DETR-S
mAP@50-9556.555.457.455.1
mAP@5073.672.375.373.5
mAP small40.837.638.435.2
FPS (mean)0.40.50.81.0
Total ms/image2507.602028.621281.03959.45
Inference ms2500.592016.361263.52944.54
Params (M)55.751.233.732.1
GFLOPs193.988.9--
Input size640640576512
LicenseAGPL-3.0non-permissiveApache-2.0Apache-2.0
YOLO26-X vs YOLO-NAS-M vs RF-DETR-M vs RF-DETR-S on Raspberry Pi 5, PyTorch FP32, batch 1. mAP shown in percent form.

Medium: RF-DETR-N covers six restricted models

RF-DETR-N reaches 51.4 mAP at 2.1 FPS. YOLO11-L reaches 52.3 at 0.7 FPS, YOLOv8-L 52.0 at 0.6, YOLOv5lu 51.5 at 0.6, and YOLO-NAS-S 51.8 at 0.9. It is also 0.8 points ahead of YOLO11-M and 2.0 ahead of YOLOv8-M, which run 0.9 and 1.1 FPS. On this board the Apache-2.0 pick is the faster one by a wide margin.

MetricYOLO11-LYOLOv8-LYOLO-NAS-SRF-DETR-N
mAP@50-9552.352.051.851.4
mAP@5068.768.669.069.9
mAP small35.434.731.227.8
FPS (mean)0.70.60.92.1
Total ms/image1385.271762.921049.54479.57
Inference ms1377.031754.541037.35468.97
Params (M)25.343.719.130.5
GFLOPs86.9165.232.8-
Input size640640640384
LicenseAGPL-3.0AGPL-3.0non-permissiveApache-2.0
YOLO11-L vs YOLOv8-L vs YOLO-NAS-S vs RF-DETR-N on Raspberry Pi 5, PyTorch FP32, batch 1. mAP shown in percent form.

Small and nano: close, with one speed gap

DEIMv2-N reaches 46.7 mAP at 2.4 FPS, against YOLO26-S at 47.4 and YOLO11-S at 45.7, both at 2.3 FPS. At the nano end YOLOv9-T (MIT) is 1.3 points ahead of YOLO26-N but runs 2.9 vs 5.0 FPS. PicoDet-M matches YOLOv8-N at 37.6 vs 36.7 mAP and 5.0 vs 5.3 FPS. DEIMv2-Femto beats YOLOv5nu on both axes: 34.5 vs 33.8 mAP at 7.4 vs 5.4 FPS.

MetricYOLO26-SYOLO26-NDEIMv2-NYOLOv9-T
mAP@50-9547.440.146.741.4
mAP@5063.855.664.056.2
mAP small29.519.728.821.5
FPS (mean)2.35.02.42.9
Total ms/image441.80198.10423.00346.25
Inference ms435.06191.50413.18340.05
Params (M)9.52.43.62.0
GFLOPs20.75.46.94.0
Input size640640640640
LicenseAGPL-3.0AGPL-3.0Apache-2.0MIT
YOLO26-S vs YOLO26-N vs DEIMv2-N vs YOLOv9-T on Raspberry Pi 5, PyTorch FP32, batch 1. mAP shown in percent form.

The permissive leaderboard on the Raspberry Pi 5

The permissive field runs past every restricted model at the top. DEIMv2-X leads at 61.3 mAP, then ECDet-X at 61.1, both at 0.3 FPS on this board. The table lists the permissive models with verified rows, sorted by accuracy.

#ModelmAP@50-95FPSms/imageInput (px)Params (M)License
1DEIMv2-X61.30.33711.7864051.2Apache-2.0
2ECDet-X61.10.33635.0864049.9Apache-2.0
3ECDet-L60.10.32850.2564033.0Apache-2.0
4D-FINE-L60.00.42305.1864031.2Apache-2.0
5RF-DETR-L58.60.52110.2270433.9Apache-2.0
6DEIMv2-L58.60.42802.2364032.5Apache-2.0
7ECDet-M58.40.51868.7364019.4Apache-2.0
8DEIM-L57.80.42295.5564031.2Apache-2.0
9D-FINE-M57.80.61577.2264019.6Apache-2.0
10RT-DETRv4-L57.80.42303.7764031.2Apache-2.0
11RF-DETR-M57.40.81281.0357633.7Apache-2.0
12RT-DETRv4-M56.50.61588.4564019.6Apache-2.0
13YOLOv9-C56.40.52153.8064025.5MIT
14DEIMv2-M56.00.51913.1964018.4Apache-2.0
15RT-DETR-R5055.90.33317.5764042.9Apache-2.0
Ranked by mAP@50-95 on Raspberry Pi 5, PyTorch FP32, batch 1. Permissive licenses only.

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.