Ultralytics YOLO26: Unified Real-Time End-to-End Vision Models

Ultralytics YOLO26 is a unified real-time vision model family that improves upon previous YOLO detectors by eliminating non-maximum suppression, removing Distribution Focal Loss, and introducing advanced training techniques including a hybrid Muon-SGD optimizer and a label assignment strategy that ensures detection of small objects. The model family spans five scales and supports multiple tasks—detection, instance segmentation, pose estimation, classification, and oriented detection—achieving 40.9-57.5 mAP on COCO with latencies of 1.7-11.8 ms on T4 TensorRT hardware. An open-vocabulary extension called YOLOE-26 enables text- and prompt-free inference for expanded capabilities.