YOLO12 (n, s, m, l, x) COCO weights for MLX
Official Ultralytics YOLO12 COCO weights, converted to NumPy .npz for use with yolo12-mlx, a YOLO12 implementation on Apple MLX.
These are the original weights, unchanged: only the tensor layout is converted (convolution kernels go from OIHW to OHWI, MLX's layout). No retraining, no fine-tuning.
| File | Scale | Parameters |
|---|---|---|
yolo12n-coco.npz |
n | 2.6 M |
yolo12s-coco.npz |
s | 9.3 M |
yolo12m-coco.npz |
m | 20.3 M |
yolo12l-coco.npz |
l | 26.6 M |
yolo12x-coco.npz |
x | 59.4 M |
Usage
git clone https://github.com/youakrim/yolo12-mlx && cd yolo12-mlx
scripts/setup.sh
python -m yolo12mlx.predict yolo12s-coco.npz image.jpg --scale s --nc 80
To fine-tune on your own classes, pass the file to python -m yolo12mlx.train --weights yolo12s-coco.npz --scale s --nc <classes> .... Layers whose shape depends on the number of classes are skipped and reinitialised.
Fidelity
Checked against the Ultralytics PyTorch models on one image at 640 px: outputs agree to within 0.011 on class scores and 3 px on boxes for all five scales. COCO mAP has not been re-measured on the MLX port.
License and attribution
AGPL-3.0, inherited from Ultralytics YOLO12, from which these weights are derived. See LICENSE. YOLO12 is described in YOLOv12: Attention-Centric Real-Time Object Detectors (Tian et al.).