Ultralytics YOLO26: Real-time end-to-end vision models that drop NMS and DFL

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

Ultralytics published YOLO26, a real-time vision model family that removes NMS and DFL for true end-to-end inference. It uses a dual-head design, a MuSGD optimizer borrowed from LLM training, and a label assignment strategy that guarantees small-object coverage. Five scales (n/s/m/l/x) hit 40.9–57.5 mAP on COCO at 1.7–11.8 ms on T4 TensorRT. An open-vocabulary variant, YOLOE-26, reaches 40.6 AP on LVIS minival with text prompts. Code and models are open-sourced.

Why it matters: YOLO26 drops NMS and DFL for a clean end-to-end pipeline — a real architectural change with concrete technical hooks (dual-head design, MuSGD optimizer). Two drags on the score: it's a paper release with no product or pricing, and pure vision detection papers sit at the edge o...