YOLOv5: the 2022 release tag, the 2026 commits, and the exports you have to install yourself

requirements.txt pulls in the successor package as a dependency

Install the requirements and you have also installed the package the README tells you to move to. The base list ends with ultralytics>=8.4.132 and ultralytics-thop>=2.1.6 for FLOPs computation, alongside torch>=1.8.0 and torchvision>=0.9.0, with floors of numpy>=1.23.5 and pillow>=10.3.0. That coupling is worth knowing before you start: this repository cannot be run in isolation from the newer ultralytics code base, so a team that wants YOLOv5 pinned and frozen is still tracking a dependency that moves on its own schedule. The README's ecosystem section is blunt about the direction of travel, describing the ultralytics package as the actively maintained one for the newest architectures, pose estimation, oriented object detection and a unified Python and CLI interface, and the top banner now advertises a later model generation.

Every export backend is commented out in requirements.txt

export.py sits at the top of the tree, so the export story looks covered, and then the dependency file says otherwise. The entire export section is commented: coremltools>=6.0 for CoreML, onnx>=1.10.0 with onnxslim>=0.1.82, tensorrt, scikit-learn<=1.1.2 for CoreML quantization, tensorflow>=2.0.0,<=2.19.0 with the note that it covers CPU, aarch64 and macos builds, tensorflowjs>=3.9.0, and openvino>=2024.0.0. Consequence for a reader: a documented install produces a repository whose export script cannot run, because no backend is present. You choose the backend, uncomment the line and accept its version range, and the tensorflow cap shows this is a considered range rather than an open one. Anyone planning an ONNX or TensorRT deployment should treat that dependency decision as part of the work.

A fresh clone has no augmentation, no COCO mAP and no screenshots

The extras section is commented out too, and two of those lines collide with things the README advertises. pycocotools>=2.0.6 carries the comment COCO mAP, and albumentations>=1.0.3 is the augmentation library, both absent after pip install -r requirements.txt. More concretely, mss, commented out and labelled screenshots, is what the screen source in the quickstart needs:

# Run inference on a screen capture
python detect.py --weights yolov5s.pt --source screen

So the documented command for screen capture is not runnable on a documented install until you add the package yourself. The same applies to the logging extras, where tensorboard>=2.4.1, clearml>=1.2.0 and comet_ml are all commented, even though the v6.2 release was about ClearML and Deci.ai integrations. The file does not say which scripts import which optional package, so a missing import is how you find out.

Weights and datasets arrive at runtime, from the newest release

Nothing about the model files is in the clone. Models are downloaded automatically from the latest YOLOv5 release, and the README says datasets are fetched the same way for training, with the training section framed as reproducing COCO results. The quickstart makes the options explicit, yolov5n, yolov5s, yolov5m, yolov5l and yolov5x, loaded through PyTorch Hub:

import torch

# Load a YOLOv5 model (options: yolov5n, yolov5s, yolov5m, yolov5l, yolov5x)
model = torch.hub.load("ultralytics/yolov5", "yolov5s")  # Default: yolov5s

Two consequences. An air-gapped machine cannot complete a first run without pre-staged weights and datasets, and the weights you get are whatever the current release tag points to rather than anything your checkout identifies, so the same commit can serve different files over time. detect.py behaves the same way, downloading yolov5s.pt on demand and writing output under runs/detect.

The install is a clone, and the wheel is the side door

Two install paths exist and they give you different things. The documented one is a repository clone with dependencies from requirements.txt:

# Clone the YOLOv5 repository
git clone https://github.com/ultralytics/yolov5

# Navigate to the cloned directory
cd yolov5

# Install required packages
pip install -r requirements.txt

The other is PyTorch Hub, which gives you a model object and not the scripts. The packaging file is explicit about the priority: a comment at the top of pyproject.toml says most users run YOLOv5 directly from a repository clone with dependencies from requirements.txt, and the build backend is plain setuptools with wheel. It declares version 7.0.0, name YOLOv5, requires-python >=3.8 and a Beta development status classifier. Practically, that means pip installing the package name is not the path the project asks you to take, and anything you import in your own code comes from a clone you have to keep in step yourself.

v7.0 is the newest tag and it is from November 2022

The release record and the commit record have drifted apart by years. Published releases are v7.0 on 2022-11-22 for realtime instance segmentation, v6.2 on 2022-08-17 for classification models, Apple M1, reproducibility and the ClearML and Deci.ai integrations, and v6.1 on 2022-02-22 for TensorRT, TensorFlow Edge TPU and OpenVINO export and inference. The last push to the repository was on 2026-09-16 and the project is not archived, so work continues on master with no release behind it. For an adopter this means a choice with no good answer: pin to v7.0 and you ship code from 2022, follow master and you ship unreviewed drift, and the version field in pyproject.toml reads 7.0.0 either way. There is no cadence left to plan a migration against.

AGPL-3.0 in three files, and an enterprise form in the README

Licensing is consistent and worth reading before anything ships. LICENSE at the top level is AGPL-3.0, pyproject.toml declares license = { "text" = "AGPL-3.0" }, and the first line of that file is a header comment carrying the AGPL-3.0 licence URL. The README then points at an Ultralytics Licensing form for requesting an Enterprise License, and says nothing about what those terms cover, so a team whose compliance process rejects AGPL has a form to fill in and no summary to evaluate. The dependency file shows the same instinct in small things: setuptools>=70.0.0 is annotated Snyk vulnerability fix, packaging is annotated as the migration path for deprecated pkg_resources packages, and a protobuf<=3.20.1 cap sits commented out next to a link to issue 8012 with no note on whether it still applies.

Two agent instruction files and a citation record sit at the top

The repository carries more than code. AGENTS.md and CLAUDE.md are both at the top level, so the project maintains instructions for automated coding agents, and their contents are not visible in the top-level listing, which means anyone automating changes here has to read both before touching anything. CITATION.cff records citation metadata, and there is a Chinese README alongside the English one in README.zh-CN.md. Alongside those, the working surface is scripts rather than a package: detect.py, train.py, val.py, export.py, benchmarks.py and hubconf.py, with classify/, segment/, data/, models/ and utils/ as directories, plus tests/ and tutorial.ipynb for people who would rather read than clone. Read the two agent files first, because the scripts assume a working environment that the requirements file only partly provides.

Editorial conclusion

Use YOLOv5 when you want detection, instance segmentation or classification from a clone today and can live with a codebase whose newest tag is v7.0 from 2022-11-22 while the last push was 2026-09-16. Do not adopt it for a new architecture: the README itself redirects pose estimation, oriented bounding boxes and the newest models to the ultralytics package, and that is also a runtime dependency of this repository. Before your first run, install the backend you need for export.py, add pycocotools for COCO mAP and mss if you intend to use the screen source, and settle the licence question, because the code is AGPL-3.0 and the only other route the README names is an enterprise licence request form.

Frequently asked questions

Official sources

1. Official documentation
2. Official README
3. Project repository
4. Release notes