Ultralytics YOLO Documentation

What these docs cover

Ultralytics maintains a popular open-source Python package for YOLO models, and these are its official docs. They cover installation, a quickstart, the modes (train, val, predict, export, track and benchmark), the tasks (detection, instance segmentation, image classification, pose estimation and oriented bounding boxes), the supported models, dataset formats, integrations and a large set of practical guides. The current model families, including YOLO11 and YOLO26, are documented side by side.

Is it worth your time?

Yes, if you need a working object detector or segmenter quickly. With a labelled dataset in YOLO format and a short YAML file describing it, one call trains a pretrained model on your data and another reports mAP on the validation split. That lets you spend your effort where it matters: collecting the right images, labelling them consistently and studying the mistakes. The guide to YOLO performance metrics explains IoU, precision, recall, mAP50 and mAP50-95 in plain terms and is worth reading even if you use another library.

What to read first

1. Quickstart, to install the package and run a pretrained model.
2. The detection task page and the dataset format guide for your labels.
3. Train and val modes, including image size, epochs and early stopping.
4. The performance metrics guide, before you report any result.
5. Export mode when you need ONNX or another runtime for deployment.

Prerequisites

Python, a basic grasp of training and validation splits, and access to a GPU for anything beyond small experiments (Google Colab is enough to start).

Licence: free and paid

The package is open source under AGPL-3.0, which suits learning, research and open projects. The AGPL has obligations if you build it into a product you distribute or offer as a service without releasing your source, and Ultralytics offers an Enterprise licence for that case. Read the licence page before using it commercially.

Common questions

Do I still need to understand detection? Yes. The package makes training easy, but choosing thresholds, splitting data without leakage and reading per-class mAP are still your job.

What can I do after it? Train, evaluate and export custom detectors and segmenters for real projects.

Best for: students and engineers building a vision project who want a proven training pipeline.