ML research

Ultralytics YOLO26 targets NMS-free end-to-end detection

2026-06-24 digest developing

  • Category: Paper
  • Status: developing
  • Sources: arXiv 2606.03748, discussion
  • Summary: Glenn Jocher and colleagues at Ultralytics posted YOLO26 (arXiv 2606.03748, dated 2026-06-02), a unified real-time vision family across five scales for detection, segmentation, pose, oriented detection, and classification. Reported changes include a dual-head design for native NMS-free end-to-end inference, removal of Distribution Focal Loss, a progressive-loss training schedule, a Muon-SGD hybrid optimizer (MuSGD), and the STAL label-assignment strategy for small objects. The authors report 40.9 to 57.5 mAP on COCO at 1.7 to 11.8 ms T4 TensorRT latency across scales, and 40.6 AP on LVIS minival for the open-vocabulary YOLOE-26 variant.
  • Why it matters: YOLO is a widely used production detection framework, and native NMS-free inference removes a common deployment and latency complication.
  • Follow-up: Independent reproduction of the COCO/LVIS numbers and the released weights and license terms.

Send feedback on this story