# YOLO26 MLX

A pure MLX reimplementation of the YOLO26 object detection model for Apple Silicon, with no PyTorch dependency at runtime, covering detection, tracking, segmentation, and pose.

YOLO26 MLX re-implements the full inference and training pipeline of YOLO26, the latest generation of the Ultralytics YOLO real-time object detection family, natively in Apple's MLX framework for Metal GPU acceleration, with no PyTorch dependency at runtime. YOLO26 itself features NMS-free end-to-end detection and simplified DFL-free box regression.

The project reports COCO val2017 accuracy within about 0.2% of the official results across most model sizes, and extends beyond plain detection with multi-object tracking using ByteTrack and BoT-SORT with pure-MLX Kalman filters and MOT17 evaluation support, an instance-segmentation head matching official mask mAP, and a pose-estimation head matching official keypoint mAP. Training uses the MuSGD and AdamW optimizers with EMA, warmup, and learning-rate scheduling, and the project includes benchmarking scripts for inference and training across all of these tasks.

Website: https://github.com/thewebAI/yolo-mlx