Install Ultralytics

Get a working environment in one command — and verify GPU access.

Before we touch a model, we need a Python environment with the ultralytics package installed and ideally a GPU it can talk to. The Ultralytics quickstart covers the whole flow — pip, Conda, or Docker — and this is the most boring lesson in the course and also the one that decides whether everything else works on the first try.

Outcome

Install Ultralytics in a clean environment and confirm yolo runs end-to-end.

Fast Track

If you already know your way around, here's the short version.

Create a fresh virtual env with Python 3.9+.
pip install ultralytics.
yolo checks — read the output, fix anything red.

Hands-on

Link to this sectionPick a Python#

PyTorch and Ultralytics installation

Use a recent Python — 3.9 to 3.12 are well supported. The cleanest setup is a fresh virtual environment so the install doesn't fight your other projects:

python -m venv .venv
source .venv/bin/activate            # Linux/macOS
# .venv\Scripts\activate             # Windows PowerShell

python --version                     # confirm 3.9+

Link to this sectionInstall#

The Ultralytics quickstart starts with the package on PyPI and pulls in PyTorch automatically:

pip install ultralytics

That installs:

The ultralytics Python package.
The yolo command-line tool.
A pinned PyTorch matching your platform.

Use a GPU when you have one

Training on CPU is fine for tiny experiments but painful for anything real. If you have an NVIDIA GPU, install a CUDA-enabled PyTorch first; pip will then keep the GPU build instead of overwriting it. On Apple Silicon, MPS works out of the box.

Link to this sectionVerify#

The most useful first command after install is the same sanity check covered in the quickstart:

yolo checks

It prints platform, Python version, PyTorch version, CUDA / MPS availability, and the version of ultralytics itself. Any red text is a setup issue you want to fix now, not in lesson 6.

A healthy output looks like this:

Ultralytics 8.x.x 🚀 Python-3.11.7 torch-2.5.0+cu121 CUDA:0 (NVIDIA RTX A4000, 16 GB)
Setup complete ✅ (8 CPUs, 32 GB RAM, 200 GB disk)

If you don't see your GPU, you don't need it for this course — but every epochs=100 will take 10× longer.

Link to this sectionA first prediction#

End the install verification with an actual prediction using the latest YOLO26 nano weights. If this works, the rest of the course works:

yolo predict model=yolo26n.pt source='https://ultralytics.com/images/bus.jpg'

You should see a downloaded model, an inference run, and a saved annotated image under runs/detect/predict/. Every default in that command — model, source, confidence, image size — is documented in the configuration reference.

Link to this sectionCommon install gotchas#

Symptom	Fix
`ImportError: numpy._core.multiarray failed`	Reinstall numpy / pin torch and ultralytics together: `pip install --upgrade --force-reinstall ultralytics`
CUDA not detected	Match torch wheel to CUDA: see pytorch.org/get-started and reinstall before `ultralytics`.
Permission errors writing `runs/`	You're in a read-only dir. `cd` into your project directory first.
`yolo: command not found`	Activate the venv (`source .venv/bin/activate`). Pip installs into the venv's bin/scripts dir.

If you hit something exotic, the Ultralytics GitHub repo is the canonical source — search closed issues before opening a new one.

Try It

Run yolo checks and yolo predict model=yolo26n.pt source='https://ultralytics.com/images/bus.jpg' end-to-end. Both should complete without errors and the prediction should save an annotated image.

Done When

You've finished the lesson when all of these are true.

yolo checks runs and the output is healthy (no red errors).
yolo predict ... saves an annotated image to runs/detect/predict/.
You know whether you're training on GPU, MPS, or CPU.

Show solution

yolo checks
yolo predict model=yolo26n.pt source='https://ultralytics.com/images/bus.jpg'

What's next

Now that the environment works, let's actually look at what predictions look like in Python.