On Jetson Orin, one MIT model sits within 0.6 mAP points of the largest restricted models and runs faster than all of them. YOLOv9-M (MIT) reaches 56.1 mAP at 8.0 FPS. YOLO26-X (AGPL-3.0) reaches 56.5 at 5.2 FPS, and YOLO-NAS-L reaches 56.3 at 5.0. At the small end the trade flips: the permissive match keeps the accuracy but runs slower than YOLO11-S or YOLO26-S.
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 23 restricted models and 55 permissive ones on the same Jetson Orin protocol. 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: YOLOv9-M covers YOLO26-X and YOLO-NAS-L
YOLOv9-M reaches 56.1 mAP at 8.0 FPS. YOLO26-X sits 0.4 points higher at 56.5 mAP but runs 5.2 FPS. YOLO-NAS-L lands at 56.3 mAP and 5.0 FPS, and YOLO-NAS-M at 55.5 mAP and 6.6 FPS. One MIT model sits within 0.6 points of all three and runs faster than each.
| Metric | YOLO26-X | YOLO-NAS-L | YOLO-NAS-M | YOLOv9-M |
|---|---|---|---|---|
| mAP@50-95 | 56.5 | 56.3 | 55.5 | 56.1 |
| mAP@50 | 73.6 | 73.3 | 72.3 | 72.7 |
| mAP small | 40.8 | 36.1 | 37.6 | 37.2 |
| FPS (mean) | 5.2 | 5.0 | 6.6 | 8.0 |
| Total ms/image | 193.65 | 200.17 | 152.02 | 124.84 |
| Inference ms | 181.78 | 180.52 | 133.21 | 114.28 |
| Peak VRAM (MB) | 386 | 470 | 365 | 183 |
| Params (M) | 55.7 | 67.0 | 51.2 | 20.1 |
| GFLOPs | 193.9 | 116.6 | 88.9 | 38.7 |
| Input size | 640 | 640 | 640 | 640 |
| License | AGPL-3.0 | non-permissive | non-permissive | MIT |
Medium: D-FINE-S against the large Ultralytics models
D-FINE-S (Apache-2.0) reaches 53.4 mAP at 10.3 FPS. YOLO26-L reaches 53.8 at 9.3 FPS. YOLO11-X reaches 53.6 at 5.1 FPS, and YOLOv8-X 53.0 at 5.4 FPS. D-FINE-S is also 1.6 points ahead of YOLO-NAS-S, 53.4 vs 51.8 mAP, at 10.3 vs 10.2 FPS. In this band the permissive pick gives up at most 0.4 points and is the fastest of the group.
| Metric | YOLO26-L | YOLO11-X | YOLO-NAS-S | D-FINE-S |
|---|---|---|---|---|
| mAP@50-95 | 53.8 | 53.6 | 51.8 | 53.4 |
| mAP@50 | 70.7 | 70.3 | 69.0 | 69.8 |
| mAP small | 37.2 | 37.1 | 31.2 | 37.9 |
| FPS (mean) | 9.3 | 5.1 | 10.2 | 10.3 |
| Total ms/image | 107.24 | 196.09 | 98.00 | 96.79 |
| Inference ms | 97.69 | 180.75 | 75.99 | 85.00 |
| Peak VRAM (MB) | 222 | 391 | 173 | 139 |
| Params (M) | 24.8 | 56.9 | 19.1 | 10.3 |
| GFLOPs | 86.4 | 194.9 | 32.8 | 25.0 |
| Input size | 640 | 640 | 640 | 640 |
| License | AGPL-3.0 | AGPL-3.0 | non-permissive | Apache-2.0 |
Small: accuracy holds, speed does not
D-FINE-N matches YOLO11-S on accuracy, 45.8 vs 45.7 mAP, but runs 11.7 FPS against 20.7. DEIM-N sits 0.6 points under YOLO26-S at 11.3 vs 20.1 FPS. YOLOX-S matches YOLOv8-S at 44.3 vs 44.2 mAP, at 16.9 vs 22.1 FPS. At the nano end YOLOX-Tiny leads YOLOv5nu by 1.7 points, at 20.3 vs 26.1 FPS. If your Jetson budget is frame rate on a small model, staying permissive costs throughput in PyTorch.
| Metric | YOLO11-S | YOLO26-S | D-FINE-N | DEIM-N |
|---|---|---|---|---|
| mAP@50-95 | 45.7 | 47.4 | 45.8 | 46.8 |
| mAP@50 | 62.3 | 63.8 | 62.8 | 63.8 |
| mAP small | 27.8 | 29.5 | 25.8 | 28.3 |
| FPS (mean) | 20.7 | 20.1 | 11.7 | 11.3 |
| Total ms/image | 48.26 | 49.88 | 85.52 | 88.46 |
| Inference ms | 36.50 | 40.47 | 73.84 | 76.67 |
| Peak VRAM (MB) | 110 | 353 | 64 | 64 |
| Params (M) | 9.4 | 9.5 | 3.8 | 3.8 |
| GFLOPs | 21.5 | 20.7 | 7.0 | 7.0 |
| Input size | 640 | 640 | 640 | 640 |
| License | AGPL-3.0 | AGPL-3.0 | Apache-2.0 | Apache-2.0 |
The permissive leaderboard on Jetson Orin
The permissive field runs past every restricted model at the top. D-FINE-X leads at 61.4 mAP, then DEIMv2-X at 61.3 and ECDet-X at 61.1. The table lists the permissive models with verified rows on this device, sorted by accuracy.
| # | Model | mAP@50-95 | FPS | ms/image | Input (px) | Params (M) | License |
|---|---|---|---|---|---|---|---|
| 1 | D-FINE-X | 61.4 | 2.9 | 350.00 | 640 | 62.6 | Apache-2.0 |
| 2 | DEIMv2-X | 61.3 | 2.7 | 370.10 | 640 | 51.2 | Apache-2.0 |
| 3 | ECDet-X | 61.1 | 2.9 | 349.46 | 640 | 49.9 | Apache-2.0 |
| 4 | ECDet-L | 60.1 | 3.4 | 295.97 | 640 | 33.0 | Apache-2.0 |
| 5 | RT-DETRv4-X | 60.0 | 2.8 | 351.81 | 640 | 62.6 | Apache-2.0 |
| 6 | D-FINE-L | 60.0 | 4.7 | 211.67 | 640 | 31.2 | Apache-2.0 |
| 7 | DEIM-X | 59.6 | 2.9 | 349.81 | 640 | 62.6 | Apache-2.0 |
| 8 | DEIMv2-L | 58.6 | 3.3 | 303.84 | 640 | 32.5 | Apache-2.0 |
| 9 | RF-DETR-L | 58.6 | 4.0 | 248.49 | 704 | 33.9 | Apache-2.0 |
| 10 | ECDet-M | 58.4 | 5.0 | 200.81 | 640 | 19.4 | Apache-2.0 |
| 11 | RT-DETR-X | 57.9 | 3.0 | 338.14 | 640 | 67.4 | Apache-2.0 |
| 12 | DEIM-L | 57.8 | 4.8 | 209.33 | 640 | 31.2 | Apache-2.0 |
| 13 | D-FINE-M | 57.8 | 7.0 | 143.31 | 640 | 19.6 | Apache-2.0 |
| 14 | RT-DETRv4-L | 57.8 | 4.7 | 210.80 | 640 | 31.2 | Apache-2.0 |
| 15 | RF-DETR-M | 57.4 | 5.7 | 174.47 | 576 | 33.7 | Apache-2.0 |
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.
