Verdict

For 30 FPS on a Raspberry Pi 5 with a Hailo-8, use YOLOv9-M: 53.3 mAP at 30.1 FPS, MIT license. YOLOv9-C is the most accurate model here at 54.8 mAP but runs at 21.2 FPS. Above 60 FPS, YOLOX-S gives 41.1 mAP at 95.8 FPS. Every row runs INT8 on the NPU, so the mAP already includes the quantization loss.

Fastest models on the Hailo-8

#ModelmAP@50-95FPSms/imageInput (px)Params (M)License
1YOLOX-Tiny33.5195.65.114165.1Apache-2.0
2YOLOX-S41.195.810.446409.0Apache-2.0
3YOLOv9-T25.776.513.086402.0MIT
4YOLO-NAS-S44.539.725.2064019.0non-permissive
5YOLOv9-S32.036.327.526407.2MIT
6YOLOX-M43.734.828.7664025.3Apache-2.0
7YOLOv9-M53.330.133.2364020.1MIT
8YOLO-NAS-M47.421.746.0564051.1non-permissive
9YOLOv9-C54.821.247.0864025.5MIT
10YOLO-NAS-L52.415.464.8964067.0non-permissive
All 10 measured detectors on Raspberry Pi 5 + Hailo-8, INT8, batch 1, by throughput.

Two rows look broken. YOLOv9-T and YOLOv9-S score far below the same models in FP32 on the Pi CPU, and below YOLOX-Tiny. YOLOv9-M and YOLOv9-C barely lose anything. Treat T and S as a bad INT8 compile, not as what those models can do on a Hailo-8, until they are rerun.

YOLOX-Tiny is the fastest row at 195.6 FPS, but it runs at 416 px. Every other model here runs at 640 px.

Most accurate models

#ModelmAP@50-95FPSms/imageInput (px)Params (M)License
1YOLOv9-C54.821.247.0864025.5MIT
2YOLOv9-M53.330.133.2364020.1MIT
3YOLO-NAS-L52.415.464.8964067.0non-permissive
4YOLO-NAS-M47.421.746.0564051.1non-permissive
5YOLO-NAS-S44.539.725.2064019.0non-permissive
6YOLOX-M43.734.828.7664025.3Apache-2.0
7YOLOX-S41.195.810.446409.0Apache-2.0
8YOLOX-Tiny33.5195.65.114165.1Apache-2.0
9YOLOv9-S32.036.327.526407.2MIT
10YOLOv9-T25.776.513.086402.0MIT
All 10 measured detectors on Raspberry Pi 5 + Hailo-8, INT8, batch 1, by mAP@50-95.

YOLO-NAS-L reaches 52.4 mAP at 15.4 FPS. YOLOv9-M beats it on both accuracy and speed, and YOLO-NAS weights come with a license that restricts commercial use. Without the Hailo-8, the same Pi tops out at 30 FPS with a much weaker model: see the CPU-only Pi 5 guide.

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.