The Curious AI Guy

Guide

Hugging Face, in plain English

Most of the open AI models on this site live on Hugging Face. Here is what it is, what you can do there, and how to use it without getting burned.

What is Hugging Face?

Think of it as GitHub for AI models: a website where people and companies publish open AI models, datasets and demo apps, and where anyone can find and download them.

A big public libraryAccording to its own docs, the Hub hosts over 2 million models, 1.5 million datasets and 1.5 million apps.
Built like codeEvery model is a repository with version history, so you can see what changed and when.
Open to everyoneYou can browse and download public models without paying. An account is needed for some things.

The Hub is the website (huggingface.co). Hugging Face is also the company behind it, and behind popular open-source libraries such as Transformers.

The three things you will see

Models

The AI models themselves: for language, images, video, speech and more. This is what you download and run.

Example: voice models like Kokoro

Datasets

The collections of data models are trained or tested on. Mostly useful if you build or fine-tune models.

Each has a dataset card

Spaces

Small apps and demos that run in your browser, so you can try a model before you download anything.

Try before you install

How to read a model page

Every model has a model card: a README that tells you what the model is for and where it falls short. Read it before you download.

  1. License. Shown on the page, and you can filter models by it. If it says "other", there is a custom license with a link. Read it.
  2. Task. What the model is meant for, such as text generation or speech. You can filter by it.
  3. Base model. Says if this is a fine-tune, an adapter, a quantized version or a merge of another model.
  4. Datasets used. What it was trained on, when the author lists it.
  5. Files. The model weights. Prefer .safetensors files over older formats. See "Stay safe".
  6. Limitations. The model card should describe intended uses, limits and biases.

Important: the license of the model files can differ from the license of the code on GitHub. That is why every AI entry on this site says to check the model page.

What you can do with it

Find an open model. Filter by task, language and license, and see which models are built on which others.

Try it in the browser. Many models have a demo Space or a widget, so you can test the quality first.

Download and run it yourself. Tools like ComfyUI download models from Hugging Face for you, and you can also get the files by hand.

Follow a team. Companies and research groups have organization pages with all their models in one place.

On this site: the ComfyUI guide and the voice guide both use models from Hugging Face. ComfyUI's own docs say its automatic model downloads usually come from there.

About Spaces (the demos)

As a visitor

  • You can open public Spaces and use them in your browser.
  • On free hardware, a Space goes to sleep when nobody uses it, so it can take a moment to wake up.

If you want to host one

  • Static Spaces are free for everyone.
  • Gradio and Docker Spaces need a paid plan. Free personal accounts can host up to 2 Gradio Spaces on ZeroGPU.
  • Better hardware, including GPUs, is billed by the hour.

"Gated" models: you may have to ask first

Some models ask you to agree to terms and share your contact details before you can download them. That is a gated model.

  1. Make a free account and log in. Requesting access can only be done from your browser.
  2. Open the model and agree. You share your username and email address with the authors, and sometimes answer extra questions.
  3. Wait if needed. Some models grant access at once. Others are approved by hand, which takes time.
  4. Download. In a script, you need an access token. The Hugging Face tools can log you in with the command hf auth login.
Know the rules: the authors can block your access at any time, even after approving it. Authors can also exclude users in the European Union with a setting, if their license does not allow distribution there. If you are in the EU, you may see a model you are not allowed to download.

Staying safe

The real risk

  • Some model files can run code. The common PyTorch format (pickle) can execute code when you load it. Hugging Face's own security docs say not to load them from sources you do not trust.
  • The Hub scans files, but it is not foolproof. It runs an antivirus scan and checks the imports in pickle files. The docs say it is your responsibility to decide what is safe.

What to do

  • Prefer .safetensors files. It is a simple weights format that does not run code.
  • Use models from people and organizations you trust. Look at who published it and how many people use it.
  • Read the model card and the license before you download or use it commercially.

New to open source? Start here. Want to run a model? ComfyUI or voice models.

Looking for a model to try?

Find by what you pay for Skills and repos Follow on Instagram