Unboxing the AI Stack
Organization · L02 · Language Models

Hugging Face

Hugging Face is the company behind the Hugging Face Hub, where AI labs and individuals publish open-weight models and datasets, and the Transformers library for running them. It was founded in New York City in 2016, according to Wikipedia,1 and the paper describing Transformers was first posted to arXiv on Oct 9, 2019.2 Hugging Face’s own spring 2026 report says the Hub grew in 2025 to 13 million users, more than 2 million public models and over 500,000 public datasets Vendor-stated.3 On Sep 3, 2026, Nvidia announced an agreement to acquire Hugging Face, saying the Hub would remain an open platform and that Nvidia compute would not be required to build on or deploy through it; Nvidia put the Hub at more than 3 million models and more than 18 million developers, researchers and creators Vendor-stated.4

Hub download counts underpin much of the measurement of open models. A download is any GET or HEAD request to a repository’s query file, config.json by default, and the count doesn’t remove duplicate users or automated pipelines.5 On those counts, Hugging Face’s report says Chinese models passed US models in monthly and overall downloads, and that Alibaba has more derivative models on the Hub than Google and Meta combined, as Chinese Open-Weight Models: Downloads, Derivatives and Token Share sets out.3

Hugging Face also runs benchmarks and builds models. Its Open ASR Leaderboard ranks speech-to-text models by average word error rate (WER), the share of words a transcript gets wrong; its short-form English track ranked 54 models in its Oct 9, 2026 version Measured, and What Word Error Rate Measures, and Why Speech-to-Text WERs Disagree is built on it.67 In robotics, its SmolVLA, posted to arXiv on June 2, 2025, is a 450-million-parameter vision-language-action model released under Apache 2.0 through its LeRobot project Vendor-stated, one of the models compared in How Vision-Language-Action Robot Models Work.89

Illustration: the ten-layer AI stack, with Hugging Face’s main layer, L02 Language Models, pulled out.

Key facts

Founded2016, New York City (per Wikipedia)1
Main productsHugging Face Hub; Transformers library (paper Oct 9, 2019)23
Hub size, 202513M users; 2M+ public models; 500K+ public datasets Vendor-stated3
Hub size, Sep 2026 (per Nvidia)3M+ models; 18M+ developers, researchers and creators Vendor-stated4
OwnershipNvidia announced an agreement to acquire it, Sep 3, 20264
How downloads are countedEvery GET or HEAD request to a repo’s query file (config.json by default)5
Open ASR Leaderboard54 models ranked, short-form English (Oct 9, 2026 version) Measured67
SmolVLA450M parameters; Apache 2.0; June 2025 Vendor-stated89

As of . How figures are labeled.

Appears in

Sources

  1. Wikipedia, Hugging Face (secondary source, used for the founding year and city only), retrieved 2026-10-13.
  2. Thomas Wolf et al. (Hugging Face), HuggingFace’s Transformers: State-of-the-art Natural Language Processing, arXiv:1910.03771v1 (2019-10-09), retrieved 2026-10-13.
  3. Hugging Face, State of Open Source on Hugging Face: Spring 2026, published 2026-03-17, retrieved 2026-10-13.
  4. Nvidia, NVIDIA to Acquire Hugging Face (2026-09-03), also listed in the NVIDIA Newsroom, retrieved 2026-10-13.
  5. Hugging Face, Models Download Stats (Hub documentation), retrieved 2026-10-13.
  6. Hugging Face, Open ASR Leaderboard (version 09-10-2026), retrieved 2026-10-13.
  7. Hugging Face, open-asr-leaderboard-results: english_short_latest.csv, revision d3c0e9a, retrieved 2026-10-13.
  8. Mustafa Shukor et al. (Hugging Face), SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics, arXiv:2506.01844v1 (2025-06-02), retrieved 2026-10-13.
  9. Hugging Face (LeRobot), lerobot/smolvla_base model card (Apache-2.0), retrieved 2026-10-13.