If you have downloaded an open AI model, there is a good chance it came from Hugging Face. The company runs the platform where researchers and companies publish models, datasets and demos — often described as the GitHub of machine learning — and maintains libraries used by millions of developers.

An unexpected start

Hugging Face was founded in 2016 in New York by French entrepreneurs Clément Delangue, Julien Chaumond and Thomas Wolf. Its first product was a chatbot app for teenagers, named after the 🤗 emoji. While building it, the team open-sourced the natural language processing tools they had developed — and those tools became far more popular than the app.

Transformers

In 2018–2019 the company released what became the Transformers library, providing easy access to pre-trained models such as BERT and GPT-2 in PyTorch and TensorFlow. A 2019 paper described the library, which quickly became the standard way to download, fine-tune and run Transformer models. It now supports thousands of architectures across text, vision, audio and multimodal tasks.

The Hub

The Hugging Face Hub hosts hundreds of thousands of models and datasets, plus interactive demos called Spaces. Model pages include documentation called model cards, licences and community discussions. When major labs release open-weight models, the Hub is usually where they are published, and it has become essential infrastructure for open AI — see our guide to AI communities.

Other widely used libraries include Datasets, Diffusers for image generation, PEFT for parameter-efficient fine-tuning, and tools for evaluation, training and serving.

Open science

Hugging Face coordinated BigScience, a year-long collaboration of more than a thousand researchers that produced BLOOM (2022), a 176-billion-parameter multilingual language model trained openly with documented data and governance. It also hosts leaderboards and free courses — see Hugging Face’s learning resources.

Business model

The Hub is free for public content. Revenue comes from:

  • paid accounts and enterprise plans with private repositories, security and access controls;
  • compute services — hosted inference endpoints and hardware for Spaces;
  • partnerships with cloud providers and companies.

In 2023 Hugging Face raised a funding round that valued it at about $4.5 billion, with investors including Google, Amazon, NVIDIA, Salesforce and others — unusual support from companies that compete with each other, reflecting the Hub’s neutral position.

Role and debates

Hugging Face is a strong advocate of open AI, transparency and community norms such as model cards. Hosting so many models also raises responsibilities: moderating harmful content, handling licences correctly and dealing with security risks such as malicious model files. It has developed scanning tools and policies to address them.

How people use Hugging Face day to day

  • Finding models: filter the Hub by task, language, licence and size, then read the model card for intended use and limitations.
  • Running models: load a model in a few lines of code with Transformers, or try it first in a hosted demo on Spaces.
  • Sharing work: researchers publish models and datasets alongside papers so results can be reproduced.
  • Fine-tuning: libraries such as PEFT make it practical to adapt open models on a single GPU.
  • Comparing: community leaderboards and evaluation tools help shortlist models for a task.

For many developers, Hugging Face is the first stop whenever a new open model is announced — which is exactly the network effect that makes the company valuable.

Competitors and partners

Cloud providers, model labs and developer platforms all offer model catalogues of their own, yet most also partner with Hugging Face, publishing their open models on the Hub and integrating its libraries. That position — neutral ground that competitors share — is unusual and central to its strategy.

What to watch next

  • Growth of open-weight model releases hosted on the Hub.
  • Security scanning for model files and supply-chain risks.
  • Hugging Face’s robotics and on-device AI initiatives.
  • Business model evolution around compute and enterprise.

Key takeaways

  • Hugging Face was founded in 2016 as a chatbot app and pivoted to open-source tools.
  • Its Transformers library became the standard way to use pre-trained models.
  • The Hub hosts hundreds of thousands of models, datasets and Spaces.
  • It coordinated BigScience, which produced the open BLOOM model in 2022.
  • It earns revenue from enterprise plans and compute services.

Frequently asked questions

Is Hugging Face free?

Public models, datasets and most libraries are free. Private repositories, enterprise features and compute are paid.

Where does the name come from?

From the 🤗 “hugging face” emoji, the name of its original chatbot app.

Are models on Hugging Face safe?

Most are, but anyone can upload. Check the publisher, licence, model card and security scans before using a model.