Model comparison · YOLO12 vs YOLOv9

YOLO12 vs YOLOv9

A size-tier-by-size-tier comparison of YOLO12 and YOLOv9: published COCO accuracy, parameter counts, licensing and which one to reach for. Every figure is the model authors' own, linked to its source.

CompareYOLO12
AgainstYOLOv9
23%fewer params

mAP comparison

YOLO12YOLOv9
35.541.347.052.858.5mAPNanoSmallMediumLargeExtra largeModel size tierYOLO12n: 40.6 mAP, 2.6M paramsYOLO12s: 48.0 mAP, 9.3M paramsYOLO12m: 52.5 mAP, 20.2M paramsYOLO12l: 53.7 mAP, 26.4M paramsYOLO12x: 55.2 mAP, 59.1M paramsYOLOv9t: 38.3 mAP, 2M paramsYOLOv9s: 46.8 mAP, 7.2M paramsYOLOv9m: 51.4 mAP, 20.1M paramsYOLOv9c: 53.0 mAP, 25.5M paramsYOLOv9e: 55.6 mAP, 58.1M params

YOLO12

  • Nano 40.6
  • Small 48.0
  • Medium 52.5
  • Large 53.7
  • Extra large 55.2

YOLOv9

  • Nano 38.3
  • Small 46.8
  • Medium 51.4
  • Large 53.0
  • Extra large 55.6
COCO mAP at 640px, as published by each model’s authors. The y axis spans 35.5 to 58.5 rather than starting at zero, so small differences stay visible. Points are the published checkpoints; the joining lines are for reading order only.

Accuracy and size, tier by tier

Each row pairs the equivalent rung of both size ladders, so the numbers sit side by side at comparable capacity. Every figure is the model authors’ own.

  • COCO mAP50-95 @ 640px
  • mAP: higher is better
  • Params (M): lower is better
YOLO12 and YOLOv9 compared at each size tier: published COCO mAP50-95 at 640px and parameter counts in millions.
TierYOLO12YOLOv9Δ mAP
ModelmAPParamsModelmAPParams
NanoYOLO12n40.62.6MYOLOv9t38.32.0M+2.3
SmallYOLO12s48.09.3MYOLOv9s46.87.2M+1.2
MediumYOLO12m52.520.2MYOLOv9m51.420.1M+1.1
LargeYOLO12l53.726.4MYOLOv9c53.025.5M+0.7
Extra largeYOLO12x55.259.1MYOLOv9e55.658.1M-0.4

Δ mAP is YOLO12 minus YOLOv9: positive means YOLO12 is ahead at that tier. Sources: YOLO12 docs · YOLO12 paper · YOLOv9 docs · YOLOv9 paper

Additional information

  • YOLO12 posts the higher published COCO mAP at 4 of the 5 matched size tiers, averaging 0.98 mAP ahead of YOLOv9. The gap is widest at the nano tier, where YOLO12 leads by 2.3 mAP. At the top of each range, YOLOv9 reaches 55.6 mAP against 55.2.

  • 23% fewer parameters at the nano tier. YOLOv9t reaches 38.3 mAP from 2.0M parameters against YOLO12n at 40.6 mAP from 2.6M, so the saving costs 2.3 mAP. No checkpoint from either family matches a larger one from the other outright.

    Parameter count is not latency. It is a good proxy for download size and memory, but throughput depends on your hardware, batch size and export format, so measure on your own target.

YOLO12

The attention-centric generation, between YOLO11 and YOLO26.

Released
2025 · Tian, Ye & Doermann (UB / UCAS)
License
AGPL-3.0
Framework
Ultralytics (PyTorch)
Tasks
Detect, Segment, Classify, Pose, OBB

An attention-centric YOLO: area attention and residual ELAN blocks replace some of the pure-convolution stack, buying about a point of mAP over YOLO11 at a similar size.

Download YOLO12 weights →

YOLOv9

Fixes deep-network information loss with PGI and GELAN.

Released
2024 · Academia Sinica (Wang et al.)
License
GPL-3.0
Framework
Original repo + Ultralytics (PyTorch)
Tasks
Detect, Segment

Adds Programmable Gradient Information and the GELAN backbone to preserve detail through deep layers, reaching high accuracy with very few parameters.

Download YOLOv9 weights →

YOLO12 and YOLOv9 against the whole lineage

YOLO12 and YOLOv9 against every other detector here, on one shared scale. The rest are dimmed; hover any row to bring it back.

Every YOLO detector in this section, with release year, publisher, published COCO mAP50-95 range at 640px, parameter-count range and licence.
ModelReleasedmAP rangeParams rangeLicence
YOLO26202640.9 to 57.52.4M to 55.7MAGPL-3.0
YOLO12202540.6 to 55.22.6M to 59.1MAGPL-3.0
YOLO11202439.5 to 54.72.6M to 56.9MAGPL-3.0
YOLOv10202438.5 to 54.42.3M to 29.5MAGPL-3.0
YOLOv9202438.3 to 55.62.0M to 58.1MGPL-3.0
YOLOv8202337.3 to 53.93.2M to 68.2MAGPL-3.0
YOLOv5 (v5u)202034.3 to 53.22.6M to 97.2MAGPL-3.0

Licenses here cover the weights. Running either model through the Ultralytics package brings AGPL-3.0 with it, and that can reach your own application regardless of how the checkpoint itself is licensed. Check the terms against your distribution plan before you ship.

Every matchup

YOLO26 vs YOLO12YOLO26 vs YOLO11YOLO26 vs YOLOv10YOLO26 vs YOLOv9YOLO26 vs YOLOv8YOLO26 vs YOLOv5 (v5u)YOLO12 vs YOLO11YOLO12 vs YOLOv10YOLO12 vs YOLOv9YOLO12 vs YOLOv8YOLO12 vs YOLOv5 (v5u)YOLO11 vs YOLOv10YOLO11 vs YOLOv9YOLO11 vs YOLOv8YOLO11 vs YOLOv5 (v5u)YOLOv10 vs YOLOv9YOLOv10 vs YOLOv8YOLOv10 vs YOLOv5 (v5u)YOLOv9 vs YOLOv8YOLOv9 vs YOLOv5 (v5u)YOLOv8 vs YOLOv5 (v5u)

Looking for the weights themselves? Download any of these checkpoints from the model-weights hub, or see how the trackers behind them compare in the YOLO tracker benchmarks.