ultralytics/ultralytics

What is Ultralytics?

Ultralytics is the official Python package that ships the YOLO family of computer‑vision models (YOLOv3 up to the latest YOLO‑26). It provides a single, easy‑to‑install library that lets you train, evaluate, predict and export state‑of‑the‑art object‑detection, segmentation, pose, depth, classification and tracking models.

Key capabilities

• Multiple vision tasks – detection, instance‑segmentation, semantic‑segmentation, depth estimation, image classification, pose estimation, oriented‑bbox detection, and video object tracking.
• Ready‑to‑use CLI (yolo) for one‑line inference or training without writing code.
• Python API (from ultralytics import YOLO) that integrates the same functionality into your own scripts.
• Pre‑trained models – a catalog of YOLO‑26 variants (n, s, m, l, x) for each task, trained on standard datasets (COCO, Cityscapes, ImageNet, NYU‑Depth, DOTAv1, etc.).
• Fast inference – models can be exported to ONNX or TensorRT for CPU/GPU acceleration.
• Cross‑platform installation – pip, conda, Docker, or build‑from‑source.

Quick start (CLI)

# Install the package
pip install ultralytics

# Run inference with a pre‑trained model
yolo predict model=yolo26n.pt source='https://ultralytics.com/images/bus.jpg'

The command works for any supported task; just change the model argument to a segmentation, pose, etc. model.

Quick start (Python)

from ultralytics import YOLO

# Load a model (downloaded automatically on first use)
model = YOLO('yolo26n.pt')

# Train on a custom dataset (COCO‑style yaml config)
model.train(data='coco8.yaml', epochs=100, imgsz=640, device='cpu')

# Evaluate
metrics = model.val()

# Predict on an image
results = model('path/to/image.jpg')
results[0].show()

# Export to ONNX for deployment

All CLI arguments have equivalent Python keyword arguments.

Model catalog (excerpt)

Task	Model	Size (px)	mAP / mIoU	Params (M)	FLOPs (B)
Detection (COCO)	YOLO26n	640	40.9 %	2.4	5.5
Detection (COCO)	YOLO26x	640	57.5 %	55.7	194.4
Segmentation (COCO)	YOLO26m‑seg	640	box 52.5 % / mask 44.1 %	23.6	121.7
Semantic Segmentation (Cityscapes)	YOLO26l‑sem	1024×2048	82.9 % mIoU	17.8	388.2
Depth (NYU‑Depth V2)	YOLO26s‑depth	640	– (depth metrics)	9.5	20.9
Classification (ImageNet)	YOLO26n‑cls	640	– (top‑1)	2.4	5.5
Pose	YOLO26s‑pose	640	– (AP)	9.5	20.9
Oriented‑bbox (DOTAv1)	YOLO26m‑obb	640	– (mAP)	20.4	68.4

Performance numbers (mAP, speed, etc.) are measured on the COCO validation set or the appropriate benchmark dataset, using an AWS P4d instance for GPU timings.

Installation options

• pip (default) – pip install ultralytics
• conda – conda install -c conda-forge ultralytics
• Docker – pull ultralytics/ultralytics
• Source – clone the repo and install with pip install -e .

All installations require Python ≥ 3.8 and PyTorch ≥ 1.8.

Where to get help

• Docs – https://docs.ultralytics.com (full guides, API reference, tutorials)
• GitHub Issues – for bug reports and feature requests
• Discord / Reddit / Forums – community discussion and support
• Enterprise licensing – available for commercial deployments.

---

Bottom line: Ultralytics bundles the latest YOLO models into a single, well‑documented Python library that lets developers and researchers quickly run or fine‑tune cutting‑edge vision models across many tasks, with optional export to production‑ready formats.