Hugging Face

Hugging Face is a website where you can download and upload fully trained AI programs. Such programs are called models: files that have learned from many examples, for instance to write texts or recognize images. Anyone who has trained a model themselves can publish it there. Anyone who needs one searches for it there and downloads it, often for free. There are also collections of sample data and small demo applications that can be tried out directly in the browser. Behind the platform stands a company headquartered in New York and Paris, founded in 2016.

The central address of the open AI scene

Before Hugging Face, models were scattered across university servers and private download links. Each had its own format and its own way of being used. Today, over a million models sit in one place, in a comparable form. This saves developers weeks of work, because no one has to start from scratch.

As a result, the platform has also become economically important. When Meta, Google, or Alibaba release a model for free download, it usually appears there first. Download numbers are considered in the industry to be a gauge of which model is gaining traction. Investors valued the company at around 4.5 billion dollars in 2023.

It is important to distinguish it from providers like OpenAI. Hugging Face does not primarily build its own top-tier models. It provides the infrastructure on which others' models are distributed. A comparison: not the musician, but the streaming platform.

From the Hub to a library on your own computer

The core is called the Hub and works somewhat like a file storage service on the internet. Each model has its own page with the files, a description, and licensing information. The license determines what you are allowed to do with the model, for instance whether commercial use is permitted. Users can leave comments and report bugs.

This includes a software package called Transformers. It translates between the many different models and a unified way of using them. Three lines of code are often enough to load a model and ask it a question. This standardization is the company’s actual technical contribution.

A third building block is called Spaces. There, small web applications run on Hugging Face’s servers. This way, anyone can test a model in the browser without installing anything. For more computing power, you pay by the hour, similar to other cloud providers. A large part of the revenue comes precisely from this and from corporate contracts.

Where the name shows up in news and projects

In news reports you read sentences like: The new model is available on Hugging Face starting today. This usually announces that a model may be freely downloaded. Reports about rankings, too, often refer to the platform’s leaderboards. There, models are compared using standardized tests.

Anyone who codes themselves will almost inevitably come across the name. School projects on image recognition or translation often draw on models from there. Many image generators that can be tried out online also technically run in a Space. An account is not always necessary for simple downloads.

A common misconception: that everything on Hugging Face has been vetted and is safe. The platform publishes contributions largely unfiltered. There have been cases of models with hidden malware in their files. That is why it’s wise to check the source and license before putting something to serious use.