cloudparts

Hugging Face

The main public host for open model weights, datasets and demo apps, stored as Git repositories on a chunk-deduplicating storage layer, with a pay-as-you-go router that sends API calls to partner inference providers.

What it is
Model hubs
Used for
Model routers and gateways
Type
Commercial product
Made by
Hugging Face, Inc.
License
Proprietary service; the huggingface_hub client and hf CLI, and the Transformers library, are Apache 2.0
Facts checked
13 October 2026
Website

What it is

Every model, dataset and Space on the Hub is a Git repository, with commit history, branches, diffs and pull requests. Large binaries such as Safetensors weights or Parquet files don’t live in Git itself: Git keeps a small pointer file and the bytes sit in remote storage on Amazon S3. That storage layer is now Xet, from the Seattle startup Hugging Face bought in August 2024. Git LFS, which the Hub used before, stores whole files, so adding one row to a 10 GB Parquet file meant uploading 10 GB again; Xet splits files into chunks and uploads only the chunks that changed. Xet has been the default for new accounts since 23 May 2025. The huggingface_hub Python library and its hf CLI download into a local cache that Transformers and other libraries share, and --revision pins a download to a commit, branch or tag.

Model repositories carry a model card, a README with metadata such as tasks, languages and evaluation results. Spaces host demo apps built with Gradio, static HTML or Docker, on free CPU Basic hardware (2 vCPU, 16 GB) or on GPUs billed by the hour, from $0.40 for an Nvidia T4. Inference Providers puts one OpenAI-compatible endpoint, https://router.huggingface.co/v1, in front of partners such as Cerebras, Groq, Together and Replicate. A suffix on the model id picks the route: :fastest (the default, highest tokens per second), :cheapest (lowest price per output token), :preferred (your own order) or a provider name such as :groq. Hugging Face says it adds no markup to provider rates; a PRO account ($9 a month) includes $2 of inference credits a month.

On 3 September 2026 Nvidia announced an agreement to buy Hugging Face for $12.93 billion, expected to close in the first half of 2027 subject to regulatory approval. Nvidia counted more than 18 million users sharing over 3 million models, 500,000 datasets and 1 million apps, and said the Hub would stay open to every model maker, cloud and accelerator, with no requirement to use Nvidia hardware.

Why teams pick it

  • Git semantics for model weights and datasets: history, branches, diffs and pull requests, and any download can be pinned to an exact revision
  • Chunk-level deduplication with Xet, so an updated checkpoint or an appended dataset uploads only the changed bytes instead of whole files
  • Free public hosting and 100 GB of free private storage; paid plans add private storage at $18 per TB-month, with egress and CDN included at no extra cost
  • One Hugging Face token reaches many inference providers through an OpenAI-compatible router, with no markup on provider rates

Watch out for

  • Free public storage is best effort beyond “the first few gigabytes”: Hugging Face asks that large uploads be useful to the community and sends heavy users to PRO, Team or Enterprise plans
  • The Git structure still shows: Hugging Face recommends fewer than 100,000 files per repository, 10,000 entries per folder and 200 GB per file
  • Inference Providers only serves models that a partner chooses to host, and it is pay-as-you-go beyond the monthly credits in paid plans
  • Ownership is changing: Nvidia’s purchase, announced in September 2026, is due to close in the first half of 2027

How it fits with other technologies

Download weights with the hf CLI, then call a model through Inference Providers · bash

pip install -U huggingface_hub        # the hf CLI, plus hf_xet for Xet storage
hf auth login                         # paste a token from hf.co/settings/tokens

hf download gpt2 config.json          # one file into ~/.cache/huggingface/hub
hf download HuggingFaceH4/zephyr-7b-beta
hf download bigcode/the-stack --repo-type dataset --revision v1.1

# OpenAI-compatible router; the suffix picks the provider policy
curl https://router.huggingface.co/v1/chat/completions \
  -H "Authorization: Bearer $HF_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"model": "openai/gpt-oss-120b:cheapest",
       "messages": [{"role": "user", "content": "Hello"}]}'

See it in a real stack

No architecture uses Hugging Face yet. Drop it into the playground and wire it up yourself.Open playground →
Where it came from2016–2026

Clément Delangue, Julien Chaumond and Thomas Wolf founded Hugging Face in New York in 2016, named after the hugging face emoji, to build a chatbot app for teenagers. TechCrunch’s 2017 review described a digital friend you texted for fun, not an assistant. The company later open-sourced the model behind the chatbot and turned to tools for machine learning.

Its Transformers library, described in a 2019 paper, put many pretrained Transformer architectures behind one API, backed by “a curated collection of pretrained models made by and available for the community.” In 2020 Hugging Face built the first version of the Hub on Git and Git LFS because LFS was well known and a reasonable way to start. By 2024 it held 12 PB in LFS across 280 million files and served 6 PB a day, and LFS was never meant for files that large, so in August 2024 Hugging Face bought XetHub, whose founders had built Apple’s internal ML infrastructure, to replace it.

The Hub grew from model hosting into datasets, Spaces for demo apps, GPU jobs, storage buckets and a router to partner inference providers. On 3 September 2026 Nvidia agreed to buy the company for $12.93 billion, saying Hugging Face would keep its brand and stay open to every model, cloud and accelerator.

Why it’s built this way

Git repositories for models and datasets
Versioning, history and pull requests came with Git, and Git LFS was well known enough to bootstrap the Hub in 2020.
Replace Git LFS with Xet’s chunked storage
AI files are very large and change a little at a time: with LFS, adding one row to a 10 GB Parquet file meant uploading 10 GB again, while chunk deduplication sends only the new chunks.
Free public storage, paid private storage
Hugging Face wants to give the community generous free space for public work and charges for private repositories above a free tier.
Route inference to partners
Instead of one provider’s catalog, one token and one OpenAI-compatible API reach many providers, with no markup on their rates.

Timeline

  1. 2016

    Founded in New York to build a chatbot app

  2. 2019

    Transformers library paper published (October)

  3. 2020

    First version of the Hub, built on Git and Git LFS

  4. 2024

    Buys XetHub to replace Git LFS (August)

  5. 2025

    Xet becomes the default storage for new users and organizations (23 May)

  6. 2026

    Nvidia agrees to buy Hugging Face for $12.93 billion (3 September)

Sources9 · facts checked 13 October 2026

Improve this page on GitHub