Model weights · YOLO12

Download YOLO12 weights

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. Pick a size below and download the official .pt checkpoint in one click, with published COCO accuracy, size and license all in view.

AGPL-3.0Tian, Ye & Doermann (UB / UCAS) · 2025 · Ultralytics (PyTorch)

YOLO12 checkpoints, pick a size and download

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.

  • YOLO12n
    yolo12n.pt
    COCO mAP
    40.6
    Params
    2.6M
    Size
    5.3 MB
  • YOLO12s
    yolo12s.pt
    COCO mAP
    48.0
    Params
    9.3M
    Size
    18.1 MB
  • YOLO12m
    yolo12m.pt
    COCO mAP
    52.5
    Params
    20.2M
    Size
    39 MB
  • YOLO12l
    yolo12l.pt
    COCO mAP
    53.7
    Params
    26.4M
    Size
    51.2 MB
  • YOLO12x
    yolo12x.pt
    COCO mAP
    55.2
    Params
    59.1M
    Size
    113.8 MB

Scores are COCO mAP at 640px, published by the authors. Weights host: github.com. Clicking Download verifies the file and starts it straight from the official CDN.

How to load YOLO12 weights

Install the Ultralytics package, then point the loader at the checkpoint. It downloads automatically on first use, or you can pass the local path to the file you downloaded above.

pip install ultralytics
from ultralytics import YOLO

# Downloads on first use, or pass a local path to your .pt file
model = YOLO("yolo12n.pt")
results = model("image.jpg")

Want to see the architecture? Export to ONNX and open it in the ONNX visualizer for a labelled diagram with shapes and parameter counts.

Open visualizer →

Choosing a YOLO12 checkpoint

Read off the table above: what each step up the size ladder costs and returns for this family specifically, where it stops being worth it, and what the license actually permits.

What the size ladder buys

YOLO12 spans YOLO12n at 2.6M parameters to YOLO12x at 59.1M, 22.7× the model, for 14.6 more COCO mAP (40.6 to 55.2). That works out to roughly 0.26 points per additional million parameters across the whole range, and the return is front-loaded: the early steps are much cheaper than the last one. Published figures from Tian, Ye & Doermann (UB / UCAS) at 640px, not re-measured here.

Where the ladder stops paying

The best value step is YOLO12n → YOLO12s, worth 1.10 COCO mAP points per million parameters. The worst is YOLO12l → YOLO12x at 0.05, or 24× less efficient, for 32.7M extra parameters and only 1.5 more points. If you are latency- or memory-bound, that is the step to skip; if you are accuracy-bound and the compute is free, it is the only place left to get it.

Where it sits in the catalog

Against the 8 families here that publish COCO mAP, YOLO12's best checkpoint ranks 3rd at 55.2, against 57.5 for YOLO26. Read that as a tier indicator rather than a head-to-head: the families use different size ladders, input sizes and training recipes, and each number is the authors' own. The closest alternatives for object detection are YOLO26, YOLO11 and YOLOv10.

Licensing, in practice

The weights are AGPL-3.0. AGPL-3.0 is free for open-source, research and internal use; a closed-source commercial product needs an Ultralytics Enterprise license.

About YOLO12

YOLO12 is the attention-centric step in the Ultralytics line, sitting between YOLO11 and YOLO26. Its area-attention module keeps a large receptive field without the quadratic cost of full self-attention, and R-ELAN blocks stabilise training at the larger sizes. Against YOLO11 it gains roughly a point of mAP at matched parameter counts, which is real but modest; the reason to pick it over YOLO26 is not accuracy but that YOLO12 predates the NMS-free head, so anything built around classic NMS post-processing drops straight in.

Author
Tian, Ye & Doermann (UB / UCAS)
Released
2025
Tasks
Detect, Segment, Classify, Pose, OBB
Framework
Ultralytics (PyTorch)
Input size
640px
License
AGPL-3.0
Resources
Compare: YOLO26 vs YOLOv8 vs YOLO11DocsGitHubPaperLicense

More model weights to download

Each page lists every checkpoint with accuracy, parameters, license and a one-click download.

See all 19 model families →

YOLO tracker FPS benchmarks →Get in touch →
Blog

From the blog

Tutorials, code, and notes on computer vision, deep learning, and applied AI.

View all posts →

Frequently asked questions

How do I download YOLO12 weights?
Click the Download button next to any variant in the table above; we verify the file and start it straight from the official CDN. You can also let the loader fetch it automatically on first use with `model = YOLO("yolo12n.pt")`. All 5 YOLO12 checkpoints are hosted officially.
Is YOLO12 free for commercial use?
YOLO12 is released under AGPL-3.0. That is free for open-source and research; closed-source commercial deployments need an Ultralytics Enterprise license. Always confirm against the linked license text.
Which YOLO12 model size should I use?
Start with YOLO12n, the smallest and fastest, ideal for prototyping, edge and CPU. Move up the ladder only when you need more accuracy and have the compute for it. The largest variant reaches 55.2 COCO mAP.