Parameters10.0M
GFLOPs25.0
Input Size640px
Best mAP52.1%
LicenseApache-2.0
Architecture
Type
transformer
Backbone
HGNetv2
Neck
HybridEncoder
Head
DEIM (Dense O2O + MAL)
Benchmark Results
Performance on COCO val2017 across different hardware configurations
| Hardware | Runtime | mAP@50-95 | FPS | Latency | VRAM |
|---|---|---|---|---|---|
| NVIDIA Jetson Orin Nano Super 8GB | ONNX Runtime FP32 | 52.1% | 10.4 | 96.2ms | - |
| NVIDIA Jetson Orin Nano Super 8GB | PyTorch FP32 | 52.1% | 10.2 | 98.1ms | 139 MB |
| NVIDIA Jetson Orin Nano Super 8GB | TensorRT FP16 | 52.0% | 29.9 | 33.4ms | - |
| NVIDIA Jetson Orin Nano Super 8GB | TensorRT FP32 | 52.1% | 22.2 | 45.0ms | - |
| NVIDIA RTX 5070 Ti | ONNX Runtime FP32 | 52.1% | 67.5 | 14.8ms | - |
| NVIDIA RTX 5070 Ti | PyTorch FP32 | 52.1% | 33.4 | 30.0ms | 142 MB |
| NVIDIA RTX 5070 Ti | TensorRT FP16 | 51.9% | 82.4 | 12.1ms | - |
| NVIDIA RTX 5070 Ti | TensorRT FP32 | 52.1% | 97.6 | 10.2ms | - |
Speed Breakdown(NVIDIA Jetson Orin Nano Super 8GB)
8.8ms
86.2ms
3.1ms
Preprocess
Inference
Postprocess (NMS)
Usage with LibreYOLO
from libreyolo import LibreYOLO
# Load model (auto-downloads from HuggingFace if not found locally)
model = LibreYOLO("LibreDEIMs.pt")
# Run inference
result = model("image.jpg", conf=0.25, iou=0.45)
# Process results
print(f"Found {len(result)} objects")
print(result.boxes.xyxy) # bounding boxes (N, 4)
print(result.boxes.conf) # confidence scores (N,)
print(result.boxes.cls) # class IDs (N,)detrnms-freePaper: 49% mAP