YOLO12 for MLX
Collection
YOLO12 COCO weights for Apple MLX. mAP identical to PyTorch on COCO val2017 (n 40.5, s 47.6, m 52.2, l 53.5, x 54.8). Code: github.com/youakrim/yolo12 • 5 items • Updated
YOLO12m COCO weights for MLX
Official Ultralytics YOLO12m COCO weights (20.3 M parameters), converted to NumPy .npz for yolo12-mlx, a YOLO12 implementation on Apple MLX. The weights are unchanged: only the tensor layout is converted (OIHW to OHWI). No retraining.
Other sizes: see the collection YOLO12 for MLX.
COCO val2017 (5000 images)
Same preprocessing (letterbox 640) and same NMS (conf 0.001, IoU 0.7, max 300 detections) for both backends, scored with pycocotools.
| Backend | mAP50-95 | mAP50 |
|---|---|---|
| MLX (this repo) | 0.5217 | 0.6914 |
PyTorch (official .pt) |
0.5217 | 0.6913 |
Speed (model forward only)
Apple M5, 640x640, batch 1, median of 100 runs after warm-up, GPU synchronised. NMS and I/O excluded.
| Backend | ms / image |
|---|---|
| MLX fp32 | 45.43 |
| MLX bf16 | 32.69 |
| PyTorch MPS fp32 | 49.62 |
| PyTorch MPS fp16 | 27.25 |
Usage
git clone https://github.com/youakrim/yolo12-mlx && cd yolo12-mlx
scripts/setup.sh
hf download youakrim/yolo12m-mlx yolo12m-coco.npz --local-dir .
python -m yolo12mlx.predict yolo12m-coco.npz image.jpg --scale m --nc 80
Licence
AGPL-3.0, inherited from Ultralytics YOLO12.
Dataset used to train youakrim/yolo12m-mlx
Collection including youakrim/yolo12m-mlx
Evaluation results
- mAP50-95 (MLX) on COCO val2017self-reported0.522
- mAP50 (MLX) on COCO val2017self-reported0.691