Ultralytics released YOLO26, a new family of real-time vision models that eliminates several longstanding bottlenecks in object detection. The architecture removes Distribution Focal Loss entirely and implements a dual-head design for native non-maximum suppression-free inference, addressing deployment complexity that has plagued previous YOLO variants.

YOLO26 introduces three coordinated training advances: Mu SGD (a hybrid Muon-SGD optimizer), Progressive Loss that shifts supervision toward inference-time heads, and STAL for positive label assignment on small objects. These changes produce a lighter detection head while extending support to instance segmentation, pose estimation, and oriented detection within a single pipeline.

The model family spans five scales from nano to extra-large, achieving 40.9-57.5 mAP on COCO at 1.7-11.8 ms T4 TensorRT latency. An open-vocabulary extension, YOLOE-26, enables text, visual, and prompt-free inference with YOLOE-26x reaching 40.6 AP on LVIS minival. Code and models are publicly available.

For developers building real-time computer vision systems, YOLO26 offers a compelling trade-off between accuracy and speed while simplifying deployment across diverse hardware targets.