Vision Analysis
Back to Leaderboard

YOLO-NAS-S

yolonas

one-stage detector with Neural-Architecture-Search (AutoNAC) backbone

PaperCode
Parameters19.0M
GFLOPs32.8
Input Size640px
Best mAP51.8%
Licensenon-permissive

Architecture

Type

one-stage

Backbone

Neural-Architecture-Search (AutoNAC)

Neck

PANet

Head

Decoupled (anchor-free, distribution focal loss)

Benchmark Results

Performance on COCO val2017 across different hardware configurations

HardwareRuntimemAP@50-95FPSLatencyVRAM
NVIDIA A100PyTorch FP3246.5%16.660.1ms174 MB
NVIDIA Jetson Orin Nano Super 8GBPyTorch FP3251.8%10.298.0ms173 MB
Raspberry Pi 5PyTorch FP3251.8%0.91049.5ms-
Raspberry Pi 5 + Hailo-8HailoRT INT844.5%39.725.2ms-

Speed Breakdown(NVIDIA A100)

4.6ms
19.2ms
36.3ms
Preprocess
Inference
Postprocess (NMS)

Usage with LibreYOLO

from libreyolo import LibreYOLO

# Load model (auto-downloads from HuggingFace if not found locally)
model = LibreYOLO("LibreYOLONASs.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,)
anchor-freenmsPaper: 47.5% mAP

Related Models (yolonas)

YOLO-NAS-MYOLO-NAS-L

Run any model with one line

LibreYOLO has the best catalogue of state-of-the-art detectors, all behind one MIT-licensed Python API.

from libreyolo import LibreYOLO, SAMPLE_IMAGE

# LibreYOLO has the best catalogue of state-of-the-art models.
model = LibreYOLO("LibreRFDETRl.pt")           # RF-DETR-L (transformer flagship)
results = model(SAMPLE_IMAGE, save=True)        # run inference, save the annotated image

# Swap in any other model, same one-line API (weights auto-download):
#   LibreYOLO("LibreYOLO9c.pt")      # YOLO9-C
#   LibreYOLO("LibreYOLOXx.pt")      # YOLOX-X
#   LibreYOLO("LibreDFINEx.pt")      # D-FINE-X
#   LibreYOLO("LibreRTDETRr50.pt")   # RT-DETR-R50