Parameters0.5M
GFLOPs0.8
Input Size320px
Best mAP27.5%
LicenseApache-2.0
Architecture
Type
transformer
Backbone
DINOv3-distilled ViT
Neck
HybridEncoder
Head
DEIM
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 | 27.5% | 43.9 | 22.8ms | - |
| NVIDIA Jetson Orin Nano Super 8GB | PyTorch FP32 | 27.5% | 15.4 | 64.9ms | 20 MB |
| NVIDIA Jetson Orin Nano Super 8GB | TensorRT FP16 | 27.5% | 89.1 | 11.2ms | - |
| NVIDIA Jetson Orin Nano Super 8GB | TensorRT FP32 | 27.5% | 81.4 | 12.3ms | - |
| NVIDIA RTX 5070 Ti | ONNX Runtime FP32 | 27.5% | 102.0 | 9.8ms | - |
| NVIDIA RTX 5070 Ti | PyTorch FP32 | 27.5% | 40.4 | 24.7ms | 24 MB |
| NVIDIA RTX 5070 Ti | TensorRT FP16 | 25.8% | 132.1 | 7.6ms | - |
| NVIDIA RTX 5070 Ti | TensorRT FP32 | 27.5% | 137.9 | 7.3ms | - |
Speed Breakdown(NVIDIA Jetson Orin Nano Super 8GB)
5.7ms
56.4ms
2.7ms
Preprocess
Inference
Postprocess (NMS)
Usage with LibreYOLO
from libreyolo import LibreYOLO
# Load model (auto-downloads from HuggingFace if not found locally)
model = LibreYOLO("LibreDEIMv2atto.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: 23.8% mAP