Measured, not copied from papers

LibreYOLO Benchmarks

LibreYOLO ships dozens of detection models behind one API. These are the numbers that tell you which one to pick: accuracy on COCO, transfer onto 100 real-world datasets, and speed on hardware you can actually buy.

100
real-world datasets in RF100-VL
6
hardware platforms measured
700+
verified benchmark runs
0.39
worst mAP drift from the original papers

COCO

Accuracy against size

The classic view: how much accuracy each architecture buys for its parameter count. Live from Vision Analysis, so it stays current as new runs land.

RF100-VL

What happens when you leave COCO behind?

COCO is 80 everyday categories: people, cars, dogs. Almost nobody ships that. The real question is transfer, so we fine-tuned on 100 unrelated real-world datasets and scored every one of them separately. Pills on a conveyor, chest X-rays, mahjong tiles, varroa mites, wildfire smoke.

Sports · 6
Document · 10
Aerial · 11
Medical · 13
Other · 15
Industrial · 22
Flora and Fauna · 23
RF100-VL
100 real-world datasets, 7 domains

Latency

Fast enough on what you own?

Accuracy does not change with hardware. Speed changes by two orders of magnitude between a datacentre GPU and a Raspberry Pi. Every run below is published as raw JSON.

Provenance

Why you can trust these numbers

Because they are ours, not the papers'. Every model LibreYOLO ships was re-measured from scratch on the full COCO val2017 set and checked against what its original authors published. Across every variant, the worst disagreement is 0.39 mAP. If a port had quietly broken, that number would show it.

See every variant against its original source