YOLO-NAS ships under a non-permissive license that restricts commercial use. On a Raspberry Pi 5 with a Hailo-8, permissive models match it. All numbers here are INT8 on the NPU. YOLOv9-M (MIT) beats YOLO-NAS-L by 0.9 mAP: 53.3 vs 52.4 mAP, at 30.1 vs 15.4 FPS. It clears YOLO-NAS-M by 5.9 points. YOLOX-M lands just behind YOLO-NAS-S at 43.7 vs 44.5 mAP. The license is not a reason to accept YOLO-NAS here.
License terms decide whether a model can ship in a commercial product. YOLO-NAS carries a non-permissive license from its authors. MIT and Apache-2.0 models carry no such restriction on use, modification, or redistribution. On the Hailo-8 NPU the permissive field covers the full YOLO-NAS range. These numbers are INT8, quantized for the accelerator. 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 and medium: YOLOv9-M covers both
YOLOv9-M reaches 53.3 mAP at 30.1 FPS on the Hailo-8. YOLO-NAS-L sits at 52.4 mAP at 15.4 FPS. YOLO-NAS-M drops to 47.4 mAP at 21.7 FPS. The MIT model leads YOLO-NAS-L by 0.9 mAP and YOLO-NAS-M by 5.9, and runs faster than both. One permissive model replaces two non-permissive ones.
| Metric | YOLO-NAS-L | YOLO-NAS-M | YOLOv9-M |
|---|---|---|---|
| mAP@50-95 | 5242.0 | 4735.0 | 5330.0 |
| mAP@50 | 6936.0 | 6377.0 | 6984.0 |
| mAP small | 3518.0 | 2948.0 | 3550.0 |
| FPS (mean) | 15.4 | 21.7 | 30.1 |
| Total ms/image | 64.89 | 46.05 | 33.23 |
| Inference ms | 0.00 | 0.00 | 0.00 |
| Params (M) | 67.0 | 51.1 | 20.1 |
| GFLOPs | 116.6 | 88.9 | 38.7 |
| Input size | 640 | 640 | 640 |
| License | non-permissive | non-permissive | MIT |
Small: YOLOX-M holds with YOLO-NAS-S
YOLOX-M reaches 43.7 mAP at 34.8 FPS on the Hailo-8. YOLO-NAS-S reaches 44.5 mAP at 39.7 FPS. The permissive model lands just behind at a comparable frame rate. In the small class the Apache-2.0 option is a close, license-clean substitute.
| Metric | YOLO-NAS-S | YOLOX-M |
|---|---|---|
| mAP@50-95 | 4446.0 | 4368.0 |
| mAP@50 | 6137.0 | 6725.0 |
| mAP small | 2834.0 | 2975.0 |
| FPS (mean) | 39.7 | 34.8 |
| Total ms/image | 25.20 | 28.76 |
| Inference ms | 0.00 | 0.00 |
| Params (M) | 19.0 | 25.3 |
| GFLOPs | 32.8 | 37.0 |
| Input size | 640 | 640 |
| License | non-permissive | Apache-2.0 |
The permissive leaderboard on the Hailo-8
Permissive alternatives cover the top of the field. YOLOv9-C tops the permissive leaderboard on the Hailo-8 at 54.8 mAP, above every YOLO-NAS variant measured. The table lists the permissive models with verified rows on this device, sorted by accuracy.
| # | Model | mAP@50-95 | FPS | ms/image | Params (M) | License |
|---|---|---|---|---|---|---|
| 1 | YOLOv9-C | 5477.0 | 21.2 | 47.08 | 25.5 | MIT |
| 2 | YOLOv9-M | 5330.0 | 30.1 | 33.23 | 20.1 | MIT |
| 3 | YOLOX-M | 4368.0 | 34.8 | 28.76 | 25.3 | Apache-2.0 |
| 4 | YOLOX-S | 4109.0 | 95.8 | 10.44 | 9.0 | Apache-2.0 |
| 5 | YOLOX-Tiny | 3353.0 | 195.6 | 5.11 | 5.1 | Apache-2.0 |
| 6 | YOLOv9-S | 3203.0 | 36.3 | 27.52 | 7.2 | MIT |
| 7 | YOLOv9-T | 2575.0 | 76.5 | 13.08 | 2.0 | MIT |
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.