Hugging Face

From Systems analysis Wiki
Jump to navigation Jump to search

Hugging Face, Inc. is an American company, headquartered in New York City, that holds a central position in the modern artificial intelligence (AI) ecosystem. The company provides an open-source platform, often called the "GitHub for machine learning," which hosts repositories for models, datasets, and demonstration applications[1]. The company's mission is to democratize AI by providing tools and fostering a global community for collaboration[2].

The company was founded in 2016 by French entrepreneurs Clément Delangue (CEO), Julien Chaumond (CTO), and Thomas Wolf (CSO)[3]. Evolving from a chatbot developer into a key piece of machine learning infrastructure, Hugging Face has become indispensable for researchers, developers, and large corporations worldwide. Its Series D round of August 2023 valued the company at $4.5 billion, which as of mid-2026 remains its last publicly disclosed valuation[4]. As of July 2026 the platform hosts more than 2 million models, over 500,000 datasets, and roughly 1 million applications (Spaces), and is used by more than 50,000 organizations[5]. In June 2026 the company announced that it had crossed $100 million in annual run-rate revenue while keeping the platform free for 97% of its users[6].

History and Development

Founding and Strategic Pivot (2016)

Initially, Hugging Face was founded in 2016 to create a consumer application—a chatbot aimed at a teenage audience. The company's name, derived from the "hugging face" emoji (🤗, U+1F917), was chosen to reflect the friendly and empathetic nature of the AI companion[1][7].

However, the chatbot did not gain significant popularity. This initial failure became a catalyst for a fundamental change in strategy. Instead of developing an end-user product, the founders decided to open-source the model that powered the chatbot[3]. The community's reaction revealed a huge demand for accessible tools for working with advanced natural language processing (NLP) models: the code published in October 2018 under the name pytorch-pretrained-bert—PyTorch implementations of Google's BERT—rapidly gained traction and became the seed of the future Transformers library[8].

The company made a strategic pivot, reorienting itself to create a machine learning platform with the mission of making AI technologies accessible to everyone, not just large corporations. Thus, the failure of a B2C product led to success in a B2D (Business-to-Developer) model, embedding the principles of openness and community focus into the company's DNA[3].

Key Milestones and Funding

After the strategic pivot, the company showed rapid growth.

  • 2019: The library was renamed Transformers. Initially developed for NLP, it quickly expanded to support models in computer vision and audio, becoming a de facto standard in the industry[9]. The same year the company released DistilBERT, a distilled version of BERT.
  • April 2021: Launch of the international BigScience research workshop, coordinated by Hugging Face together with several other research groups[3].
  • December 2021: Hugging Face acquired Gradio, a popular open-source library for quickly creating interactive demonstrations. Gradio had already become the main engine behind Hugging Face Spaces[10].
  • May 2022: A Series C funding round of $100 million valued the company at approximately $2 billion[11].
  • July 2022: The BigScience workshop concluded with the release of BLOOM, a multilingual model with 176 billion parameters distributed under an open Responsible AI License[12].
  • February 2023: A strategic partnership with Amazon Web Services made Hugging Face models available as building blocks for AWS customers[3].
  • August 2023: A Series D funding round of $235 million took place, raising the company's valuation to $4.5 billion. Google, Amazon, Nvidia, Salesforce, Intel, AMD, Qualcomm, IBM, and Sound Ventures participated in the round[4]. Total funding raised across all rounds is approximately $400 million[13].
  • 2024: Launch of LeRobot, an open-source library for robot learning, led by Rémi Cadène[14].
  • June 2024: Acquisition of Argilla, a Madrid-based platform for building and curating datasets and human feedback, reportedly for about $10 million[15][16].
  • August 2024: Acquisition of XetHub (Seattle), whose chunk-level deduplication technology replaced Git LFS as the storage backend of the Hub[17][18].
  • April 2025: The company acquired the French humanoid robotics startup Pollen Robotics, marking an expansion into embodied AI (embodied AI)[19][20].
  • July 2025: Announcement of Reachy Mini, a desktop open-source robot, as the platform passed the milestone of 10 million users[21].
  • December 2025: Announcement of Transformers v5, the first major release of the library in five years[8].
  • February 2026: ggml.ai, the team of Georgi Gerganov behind llama.cpp and the GGUF format, joined Hugging Face. The projects remain fully open-source and community-driven; the stated goal is seamless integration between Transformers as the model-definition layer and llama.cpp as the local-inference layer[22][23].
  • June 2026: The company reported crossing $100 million in annual run-rate revenue[6].

The Hugging Face Ecosystem

The Hugging Face ecosystem covers the entire machine learning model development lifecycle—from data preparation to deployment.

Hugging Face Hub

The core of the ecosystem is the Hugging Face Hub, a central web platform for collaboration. It includes:

  • Model Repositories: Git-based repositories for storing models, their weights, and configuration files. They provide version control for experiment reproducibility. Since the acquisition of XetHub, large files are stored using a chunk-based, deduplicating backend instead of Git LFS, so that editing a large file no longer requires re-uploading it in full[17].
  • Dataset Repositories: Similar repositories for storing and versioning datasets.
  • Spaces: An interactive environment for creating and demonstrating web applications (demos) based on models, using frameworks like Gradio, Streamlit, Docker, or static HTML.
  • Model Cards: Standardized documents describing the characteristics, limitations, and potential biases of models, which helps promote transparency[24].
  • Inference Providers: A unified API that routes requests to third-party inference providers, giving access to tens of thousands of models without additional service fees[5].
  • HuggingChat: An open chat interface built on open-weight models[5].

Transformers Library

Transformers is Hugging Face's flagship software product, providing a unified API for accessing hundreds of thousands of pre-trained models. Key features:

  • A standard for model definitions: the library aims to be the "source of truth" for model architectures. By the v5 release it covered more than 400 architectures (up from 40 in v4), with over 750,000 compatible checkpoints published on the Hub[8].
  • Scale of adoption: more than 3 million installations per day via pip and over 1.2 billion cumulative installations[8].
  • PyTorch as the sole backend: versions up to v4 supported PyTorch, TensorFlow, and Flax/JAX, but v5 sunset TensorFlow and Flax in favour of PyTorch, while compatibility with the JAX ecosystem is maintained through partner tools[25].
  • Ease of Use: Loading, fine-tuning, and using models can be done in just a few lines of code.
  • Efficiency: It provides access to a vast number of models, allowing users to avoid training them from scratch, which saves resources and reduces the carbon footprint[26].

Other Key Libraries

  • Datasets: A library for efficient access and processing of datasets using the Apache Arrow format.
  • Diffusers: A library of state-of-the-art diffusion models for image, audio, and video generation.
  • Tokenizers: A high-performance library written in Rust for text tokenization.
  • Accelerate: Simplifies distributed training across multiple GPUs/TPUs.
  • PEFT (Parameter-Efficient Fine-Tuning): A library of methods for efficiently fine-tuning large models.
  • TRL: Tools for post-training language models with reinforcement learning and preference optimization.
  • Safetensors: A safe and fast format for storing neural network weights, which has become the default standard in the ecosystem.
  • Text Generation Inference (TGI): An optimized toolkit for serving language models in production.
  • Transformers.js: Model inference directly in the browser.
  • smolagents: A minimal library for building AI agents.
  • LeRobot: An end-to-end library for robot learning[5].

Models and Datasets Released by Hugging Face

Beyond tooling, the company trains and releases its own open models and datasets. Notable examples include DistilBERT (2019), a compressed version of BERT; the SmolLM family of compact language models (SmolLM, SmolLM2, SmolLM3); the SmolVLM vision-language models; and SmolVLA, a vision-language-action model for robotics. On the data side, Hugging Face has published large open pre-training corpora such as FineWeb, FineWeb-Edu, Cosmopedia, and FineMath[27][28]. The company also maintains evaluation infrastructure, including the Open LLM Leaderboard, a widely used reference for ranking open models.

Open Robotics

Since 2024 the company has extended its open-source approach to physical AI. The LeRobot library provides datasets, pre-trained policies, and a unified interface for a wide range of hardware, from low-cost arms to humanoids[14][29]. Following the acquisition of Pollen Robotics, the product line spans the Reachy 2 research humanoid, the SO-100/SO-101 3D-printable robotic arms (from roughly $100), and Reachy Mini, a desktop robot priced from $299 for the tethered version and $449 for the wireless one, which integrates natively with the Hub and can be programmed in Python[30][21].

Business Model and Market Positioning

Hugging Face uses a freemium (open-core) business model, combining open access with commercial offerings for enterprise clients.

  • Free Tier: Offers unlimited hosting for public models, datasets, and applications, attracting millions of users; the company states that around 97% of its users pay nothing[6].
  • Revenue Sources:
    • PRO Subscription: An individual subscription ($9/month) with increased storage, inference credits, and ZeroGPU quota.
    • Team and Enterprise plans: Corporate products (from $20/user per month for Team and from $50/user per month for Enterprise) with SSO, storage regions, audit logs, granular access control, and dedicated support.
    • Paid Compute Resources: Paid GPU hardware for Spaces and dedicated Inference Endpoints (from $0.033 per hour), as well as volume-based storage pricing[31].

The company positions itself as a neutral, model-agnostic infrastructure platform, hosting competing model families side by side and building deep partnerships with major cloud providers (AWS, Google Cloud, Microsoft Azure) and hardware manufacturers such as Nvidia, Intel, AMD, and Qualcomm[32]. Delangue has argued that the current wave of investment constitutes an "LLM bubble" rather than an AI bubble, and has emphasized capital efficiency: as of late 2025 the company still held roughly half of the $400 million it had raised[13]. According to the company, close to half of the Fortune 500 now use the platform[33].

Mission to Democratize AI

A central element of Hugging Face's identity is its mission to democratize AI, which is realized through the principles of open source and open science.

A prominent embodiment of this philosophy is the BigScience research initiative. This open international workshop, coordinated by Hugging Face and funded in part by the French government, ran for about a year and brought together over 1,000 researchers from more than 70 countries and 250 institutions. Its result was the BLOOM model—a large multilingual language model (176 billion parameters) trained on the Jean Zay public supercomputer and capable of generating text in 46 natural languages and 13 programming languages[34]. Released in July 2022 under the Responsible AI License (RAIL), it permits broad research and commercial use but imposes restrictions on high-risk applications such as mass surveillance, disinformation, and algorithmic discrimination[35][36].

The same philosophy underlies later initiatives: the integration of ggml/llama.cpp aims to make local inference on consumer hardware a practical alternative to cloud APIs[22], while the robotics line seeks to lower the cost of entry into embodied AI[20].

Criticism and Security Challenges

The openness of the Hub also creates an attack surface for software supply-chain threats. In early 2024, researchers at JFrog reported roughly 100 malicious models on the platform; loading such a model could execute arbitrary code and open a reverse shell on the victim's machine. About 95% of these artifacts exploited PyTorch's pickle-based serialization format[37][38]. In response, Hugging Face introduced automated malware and pickle scanning, partnered with third-party model scanners, and promoted safetensors as a safe default weight format. The problem has nevertheless persisted at scale, and in January 2026 the platform was abused to distribute Android malware[39][40].

Documentation quality is a second concern: an academic study of the Hub found that only about two-thirds of models carry a model card, which limits transparency about training data, limitations, and intended use[41].

References

  1. ↑ 1.0 1.1 "What is Hugging Face? A Beginners Guide". 365 Data Science. [1]
  2. ↑ "What is Hugging Face?". IBM. [2]
  3. ↑ 3.0 3.1 3.2 3.3 3.4 "Hugging Face". Wikipedia. [3]
  4. ↑ 4.0 4.1 "Hugging Face — Funding Rounds & List of Investors". Tracxn. [4]
  5. ↑ 5.0 5.1 5.2 5.3 "Hugging Face – The AI community building the future". Hugging Face. [5]
  6. ↑ 6.0 6.1 6.2 Delangue, Clément. "We just crossed $100M annual run-rate…". X, 25 June 2026. [6]
  7. ↑ Dillet, Romain. "Hugging Face wants to become your artificial BFF". TechCrunch, 9 March 2017. [7]
  8. ↑ 8.0 8.1 8.2 8.3 "Transformers v5: Simple model definitions powering the AI ecosystem". Hugging Face Blog, 1 December 2025. [8]
  9. ↑ "The Transformers Library: standardizing model definitions". Hugging Face Blog. [9]
  10. ↑ "Gradio is joining Hugging Face!". Hugging Face Blog, December 2021. [10]
  11. ↑ "Hugging Face seals $100m Series C". Venture Capital Journal. [11]
  12. ↑ "BigScience Releases 176B Parameter AI Language Model BLOOM". InfoQ, July 2022. [12]
  13. ↑ 13.0 13.1 Perez, Sarah. "Hugging Face CEO says we're in an 'LLM bubble,' not an AI bubble". TechCrunch, 18 November 2025. [13]
  14. ↑ 14.0 14.1 "LeRobot: Making AI for Robotics more accessible with end-to-end learning". GitHub. [14]
  15. ↑ Vila Suero, Daniel. "Argilla is joining Hugging Face". Argilla Blog, 13 June 2024. [15]
  16. ↑ Wodecki, Ben. "Hugging Face Acquires AI Software Startup to Boost Datasets". AI Business. [16]
  17. ↑ 17.0 17.1 "XetHub is joining Hugging Face!". Hugging Face Blog, August 2024. [17]
  18. ↑ Nieva, Richard. "AI Unicorn Hugging Face Acquires A Startup To Eventually Host Hundreds Of Millions Of Models". Forbes, 8 August 2024. [18]
  19. ↑ Wiggers, Kyle. "Hugging Face buys a humanoid robotics startup". TechCrunch, 14 April 2025. [19]
  20. ↑ 20.0 20.1 Knight, Will. "An Open Source Pioneer Wants to Unleash Open Source AI Robots". Wired, 14 April 2025. [20]
  21. ↑ 21.0 21.1 "Hugging Face just launched a $299 robot that could disrupt the entire robotics industry". VentureBeat. [21]
  22. ↑ 22.0 22.1 "GGML and llama.cpp join HF to ensure the long-term progress of Local AI". Hugging Face Blog, 20 February 2026. [22]
  23. ↑ Willison, Simon. "ggml.ai joins Hugging Face to ensure the long-term progress of Local AI". 20 February 2026. [23]
  24. ↑ "Model Cards". Hugging Face Docs. [24]
  25. ↑ "Release Transformers v5". GitHub — huggingface/transformers. [25]
  26. ↑ "Transformers". Hugging Face Docs. [26]
  27. ↑ "HuggingFaceTB (Hugging Face Smol Models Research)". Hugging Face. [27]
  28. ↑ "SmolLM3: smol, multilingual, long-context reasoner". Hugging Face Blog, July 2025. [28]
  29. ↑ "LeRobot". Hugging Face Docs. [29]
  30. ↑ "Reachy Mini — The Open-Source Robot for Today's and Tomorrow's AI Builders". Hugging Face Blog. [30]
  31. ↑ "Pricing". Hugging Face. [31]
  32. ↑ "Hugging Face Business Breakdown & Founding Story". Contrary Research. [32]
  33. ↑ "Hugging Face's CEO on why companies are done renting their AI". TechCrunch, 10 July 2026. [33]
  34. ↑ Le Scao, T. et al. "BLOOM: A 176B-Parameter Open-Access Multilingual Language Model". arXiv:2211.05100. [34]
  35. ↑ "bigscience/bloom". Hugging Face. [35]
  36. ↑ "BLOOM (language model)". Wikipedia. [36]
  37. ↑ "Over 100 Malicious AI/ML Models Found on Hugging Face Platform". The Hacker News, March 2024. [37]
  38. ↑ "Top JFrog Security Research Discoveries of 2024". JFrog. [38]
  39. ↑ Fadilpašić, Sead. "Hugging Face platform hijacked to send out Android malware". TechRadar Pro, 30 January 2026. [39]
  40. ↑ "Hugging Face and ClawHub compromised with hundreds of malicious AI models". The Next Web, 2026. [40]
  41. ↑ "Anatomy of a Machine Learning Ecosystem: 2 Million Models on Hugging Face". arXiv:2508.06811. [41]
Retrieved from "https://systems-analysis.info/eng/index.php?title=Hugging_Face&oldid=181"