Skip to content

YOLOv26 (YOLO26) Object Detection Format

Overview

YOLOv26 (also known as YOLO26) is the latest evolution in the You Only Look Once (YOLO) series, engineered specifically for edge and low-power devices. It introduces a streamlined design that removes unnecessary complexity while integrating targeted innovations to deliver faster, lighter, and more accessible deployment. YOLOv26 uses the same object detection format as YOLOv8-v12, utilizing normalized coordinates in text files for seamless compatibility.

Info: YOLOv26 is currently in preview and under development. Performance numbers are preliminary and final releases will follow soon. For the latest updates, see: GitHub Repository: ultralytics/ultralytics

Key YOLOv26 Features

YOLOv26 maintains full compatibility with the YOLOv8-v12 label format while introducing several breakthrough innovations:

  • End-to-End NMS-Free Inference: Native end-to-end model producing predictions directly without non-maximum suppression, reducing latency and simplifying deployment
  • DFL Removal: Eliminates Distribution Focal Loss module for better export compatibility and broader hardware support on edge devices
  • MuSGD Optimizer: Hybrid optimizer combining SGD with Muon, inspired by Moonshot AI's Kimi K2 breakthroughs in LLM training
  • ProgLoss + STAL: Enhanced loss functions with notable improvements in small-object detection accuracy
  • 43% Faster CPU Inference: Specifically optimized for edge computing with significant CPU performance gains

Format Specification

YOLOv26 uses the identical format as YOLOv8, YOLOv9, YOLOv10, YOLOv11, and YOLOv12. Please refer to the YOLOv8 format documentation for complete format specifications, including:

  • Text file structure with normalized coordinates
  • Directory organization patterns
  • Configuration via data.yaml
  • Coordinate normalization formulas
  • Example annotations

Converting Annotations to YOLOv26 Format

Since YOLOv26 uses the same format as YOLOv8-v11, you can convert from other formats using Labelformat:

labelformat convert \
    --task object-detection \
    --input-format coco \
    --input-file dataset/annotations/instances_train.json \
    --output-format yolov26 \
    --output-folder dataset/yolov26_labels \
    --output-split train

The converted output will be fully compatible with YOLOv26 training and inference pipelines.