YOLO26s-OptiQ-6bit

Mixed-precision quantized YOLO26s for Apple Silicon via OptiQ

This is a mixed-precision quantized version of YOLO26s in MLX format, optimized with mlx-optiq for Apple Silicon inference via yolo-mlx.

Quantization Details

Property	Value
Target BPW	6.0
Achieved BPW	5.97
Layers at 4-bit	11
Layers at 8-bit	115
Original size	38.4 MB
Quantized size	8.9 MB
Compression	4.3x

Benchmark Results (COCO128)

Model	Total Detections	Avg/Image
OptiQ 6-bit	633	4.9
Original (FP32)	681	5.3

Detection delta: -48 (-7.0%) at 4.3x compression.

Usage

Requires mlx-optiq and yolo-mlx:

pip install mlx-optiq yolo-mlx

from optiq.models.yolo import load_quantized_yolo

model = load_quantized_yolo("mlx-community/YOLO26s-OptiQ-6bit")
results = model.predict("image.jpg")

How OptiQ Works

OptiQ measures each conv layer's sensitivity via KL divergence on detection outputs, then assigns optimal per-layer bit-widths using greedy knapsack optimization. Sensitive layers (detection head, feature pyramid) get 8-bit precision while robust backbone layers get 4-bit.

Article

For more details on the methodology and results, see: Not All Layers Are Equal: Mixed-Precision Quantization for Weights and KV Cache on Apple Silicon

Credits

• Quantization: mlx-optiq by Thin Signal
• Base model: YOLO26 by Ultralytics
• MLX runtime: yolo-mlx
• Framework: MLX by Apple