Ultralytics YOLO26: Unified NMS-free real-time vision models
Tags AI · OSS
Ultralytics released YOLO26, a unified real-time vision model family featuring native NMS-free end-to-end inference via dual-head design, achieving 40.9-57.5 mAP on COCO at 1.7-11.8 ms T4 TensorRT latency. The model uses MuSGD (hybrid Muon-SGD optimizer from LLM training), Progressive Loss, and STAL label assignment. YOLOE-26x reaches 40.6 AP on LVIS minival under text prompting for open-vocabulary inference. Code and models available on GitHub.
Technical significance
YOLO26's NMS-free design and use of LLM-derived optimizers (Muon) in computer vision training represents cross-pollination between language model and vision model techniques. The open-source availability with pre-trained weights lowers the barrier for production deployment of real-time vision systems.