Hugging Face for beginners

There is a sentence that describes Hugging Face well, and there is a sentence that sounds good. The good-sounding one is "GitHub for machine learning". It helps at first glance and gets in your way afterwards, because it stirs three things into one that belong apart.

Because when you are on huggingface.co you are not looking at one product, but at three:

The Hub is the platform. Models, datasets and demos live there as real Git repositories, each with a history, commits and a description page. This is the part that really does work like GitHub.

The libraries are Python packages you install: transformers, datasets, huggingface_hub and a dozen more. They run on your machine and do not need the platform - you can use a model with transformers that has never been anywhere near Hugging Face.

The services are what the company earns money with: compute time for demos, managed inference endpoints, a gateway to third-party providers, team features. Some of it is free, some of it costs from the first minute.

Anyone who mixes the three up asks the wrong questions. "Is Hugging Face GDPR compliant?" is one of them: for the libraries on your laptop the question is meaningless, for the Hub it is a question about metadata, and for the services it is a question about which data centre of which provider processes your inputs. Three different answers, because they are three different things.

Why this course exists

Most introductions stop after the download. Find a model, load a model, run a model - done. That is the part almost everyone knows, and it is the part that surprises the least.

The real astonishment comes later, and it has a name: Spaces. You push a file called app.py into a repo, wait two minutes, and there is a web application on the internet. With a URL. Reachable by strangers. Optionally with a GPU. No server, no deployment pipeline, no bill. Chapter 05 is the chapter this course exists for - everything before it is the road there, everything after it is the question of where the compute comes from once things get serious.

What you have at the end

What you need

Not much, and none of it costs money:

• Python 3.10 or newer and a terminal. No Docker, no Conda required.
• Around 3 GB of free disk space for cached models. We deliberately stay small.
• No graphics card. Every step in this course runs on CPU. Wherever a GPU turns up, it belongs to somebody else.
• A free account, but not before chapter 02. You can read the first two chapters without signing up.

What this course is not

Not a fine-tuning course. LoRA, QLoRA, TRL and the whole post-training workshop appear here only as names - they need a GPU and a weekend of their own.

Not a model guide. Which model is good for which task is a question that re-poses itself every couple of months; this course teaches you how to answer it yourself, instead of handing you a list that is stale on the day it is published.

And not a price list. A few figures appear in the text because they carry the argument - free or not free is an architectural decision. All of them are marked as of August 2026, and that is exactly how they should be read: as a snapshot, not as a fact.

The roadmap

Eight chapters, each with a result you can show. They build on one another, but not so strictly that you could not skip 04 if datasets do not interest you right now. Only 02 should not be skipped - after it you have the account and the CLI, and every following chapter needs both.

We start with the part that looks as though you already know it: reading a model page.