# YOLOv9-S

> YOLOv9-S: one-stage object detector from the YOLOv9 family. Available in LibreYOLO, the MIT-licensed open-source library (free for commercial use).

- Family: YOLOv9 (Academia Sinica)
- Parameters: 7.2M
- GFLOPs: 13.5
- Default input size: 640px
- Detection: anchor-free · nms
- Architecture: one-stage · GELAN backbone
- Weight license: MIT
- Paper-reported mAP@50-95: 46.8%
- Paper: https://arxiv.org/abs/2402.13616
- Original code: https://github.com/MultimediaTechLab/YOLO

## Benchmarks (COCO val2017)

| Hardware | Runtime | mAP@50-95 | FPS | Latency (ms) | VRAM (MB) |
|---|---|---|---|---|---|
| NVIDIA Jetson Orin Nano Super 8GB | ONNX Runtime FP32 | 50.4 | 12.5 | 79.8 | - |
| NVIDIA Jetson Orin Nano Super 8GB | PyTorch FP32 | 50.5 | 9.9 | 101.2 | 91 |
| NVIDIA Jetson Orin Nano Super 8GB | TensorRT FP16 | 50.4 | 33.2 | 30.1 | - |
| NVIDIA Jetson Orin Nano Super 8GB | TensorRT FP32 | 50.5 | 21.6 | 46.2 | - |
| NVIDIA RTX 5070 Ti | ONNX Runtime FP32 | 50.4 | 54.5 | 18.3 | - |
| NVIDIA RTX 5070 Ti | PyTorch FP32 | 50.5 | 30.8 | 32.5 | 92 |
| NVIDIA RTX 5070 Ti | TensorRT FP16 | 50.5 | 63.8 | 15.7 | - |
| NVIDIA RTX 5070 Ti | TensorRT FP32 | 50.5 | 75.1 | 13.3 | - |

## Usage with LibreYOLO

```python
from libreyolo import LibreYOLO

model = LibreYOLO("LibreYOLO9s.pt")
result = model("image.jpg", conf=0.25, iou=0.45)
```

Source: https://www.visionanalysis.org/model/yolov9s. Benchmarks produced with LibreYOLO (https://github.com/Libre-YOLO/libreyolo).