CtrlK
Hugging Face is an artificial intelligence company and open-source platform widely regarded as the "GitHub of machine learning." Headquartered in New York City with offices in Paris, the company provides tools, infrastructure, and a collaborative community hub that enable researchers, data scientists, and engineers to build, train, share, and deploy machine learning (ML) models. Its platform hosts millions of pre-trained models, datasets, and interactive demo applications spanning natural language processing (NLP), computer vision, speech recognition, reinforcement learning, and other AI domains. [1][2]
Technological leaders including Microsoft, Google, Meta, Apple, Amazon Web Services (AWS), Nvidia, and hundreds of other organizations rely on Hugging Face's models, datasets, and libraries. Over 30% of the Fortune 500 maintain verified accounts on the platform. [3][4] The company's mission centers on democratizing AI by making cutting-edge tools accessible to everyone, from individual hobbyists to large enterprises.
Hugging Face was founded in 2016 by three French entrepreneurs: Clement Delangue (CEO), Julien Chaumond (CTO), and Thomas Wolf (CSO). The company was originally created in New York City as a consumer technology startup building a chatbot application targeted at teenagers, serving as a kind of "AI best friend." The company's name comes from the hugging face emoji. [5][6]
Delangue brought experience in growth and marketing, while Chaumond and Wolf contributed engineering expertise and deep knowledge of computational linguistics. During this early period, the team built sophisticated natural language processing capabilities to power their chatbot's conversational abilities. [6]
The turning point for Hugging Face came in late 2018, when Google released BERT (Bidirectional Encoder Representations from Transformers), a groundbreaking language model. The Hugging Face team rapidly produced and open-sourced a PyTorch implementation of BERT within a single week. This effort attracted significant attention from the ML research community. [5]
Chaumond later noted that this moment clarified the company's strategic direction. In 2019, Hugging Face formally pivoted away from the consumer chatbot product and toward building open-source machine learning infrastructure. The team open-sourced the internal NLP tools they had developed for their chatbot and launched the first version of the Hugging Face Transformers library, which quickly became the most widely used library for working with transformer models. [5][6]
Following the pivot, Hugging Face expanded rapidly. The company launched the Hugging Face Hub as a centralized repository for models and datasets, modeled after GitHub's approach to code hosting. The platform introduced Spaces for hosting ML demos, integrated with major cloud providers, and grew its community of contributors and users. [2]
In December 2021, Hugging Face acquired Gradio, an open-source Python library for building interactive ML demos, founded by Abubakar Abid during his PhD at Stanford University. The acquisition brought Gradio's team of five engineers into Hugging Face and provided the foundation for Spaces, the platform's demo hosting feature. Since the acquisition, Gradio has grown to over 2 million monthly users and powers more than 470,000 applications. [7]
By 2023, Hugging Face had established itself as the default platform for sharing and discovering open-source AI models. The company raised a landmark $235 million Series D round that valued it at $4.5 billion, with participation from major technology companies across the industry. [8]
In February 2026, Hugging Face announced that ggml.ai, the organization behind llama.cpp, would join the company. Georgi Gerganov and the founding ggml.ai team became full-time Hugging Face employees, bringing together the model distribution layer (the Hub), model definition layer (Transformers), and local inference layer (llama.cpp) under a single organization. The ggml and llama.cpp projects remain fully open-source and community-driven, with Gerganov retaining full autonomy over technical decisions. [9]
Hugging Face has raised approximately $400 million across multiple funding rounds since its founding.
| Round | Date | Amount | Lead Investor(s) | Notable Participants |
|---|---|---|---|---|
| Seed | October 2016 | Undisclosed | The Chernin Group | Early angels |
| Series A | December 2019 | $15 million | Lux Capital | A.Capital, Betaworks, Richard Socher, Greg Brockman |
| Series B | March 2021 | $40 million | Addition VC | Lux Capital, A.Capital, Betaworks, Kevin Durant |
| Series C | April 2022 | $100 million | Sequoia Capital | AIX Ventures, Coatue, others |
| Series D | August 2023 | $235 million | Salesforce Ventures | Google, Amazon, Nvidia, AMD, Intel, Qualcomm, IBM, Sound Ventures |
The Series D valuation of $4.5 billion represented a doubling from the company's Series C valuation and was reportedly more than 100 times Hugging Face's annualized revenue at the time. [8]
Revenue has grown rapidly, from approximately $10 million in 2021 to $15 million in 2022, $70 million in 2023, and $130 million in 2024. The company's revenue growth of 367% year-over-year in 2023 was driven primarily by enterprise consulting contracts with organizations like Nvidia, Amazon, and Microsoft. Revenue streams include freemium subscriptions, API usage fees, enterprise contracts, and consulting services. [12][13]
The Hugging Face Hub is the central platform through which the company delivers its services. It functions as a collaborative, Git-based hosting platform for ML models, datasets, and demo applications. Each artifact on the Hub is stored as a Git repository, enabling versioning, branching, collaboration, and discoverability. [2]
As of early 2026, the Hub hosts over 2 million public models, more than 500,000 public datasets, and approximately 1 million demo applications (Spaces). The platform serves over 13 million users. [14][15]
The growth trajectory has been striking. The first million model repositories took over 1,000 days to accumulate starting from March 2022, while the second million arrived in just 335 days. The platform processes roughly 15 million new downloads daily, and approximately 10,000 new models are uploaded each week. [14]
Creating a new model on the platform generates a Git repository for the files associated with that ML model. Users can specify the type of open-source license, define the model's visibility (public or private), and configure metadata including the datasets used for training and the Spaces that use the model. [2]
Each model page on the Hub includes several elements:
| Element | Description |
|---|---|
| Name and tags | The model name, number of likes, and associated tags for discoverability |
| Model card | An overview of the model with documentation, code snippets, and usage instructions |
| Training and deployment | Options to train, fine-tune, or deploy the model through cloud providers |
| Metadata | Information about training datasets, Spaces using the model, and related resources |
| Files and versions | Git-based file browser showing model weights, configs, and version history |
| Community | Discussion tab for questions, feedback, and collaboration |
Datasets on the Hub are used for model training and fine-tuning and are available in multiple languages. When creating a new dataset, users name it and choose a license type. Dataset pages include a title, tags, table of contents, an embedded data preview, quick links to the GitHub repository, code snippets for loading the data through the Datasets library, and metadata about the origin, size, and models trained on the dataset. [2][3]
Spaces is the Hub's feature for hosting interactive ML demo applications. Users can build and deploy demos using Gradio, Docker, or static HTML. By default, Spaces run on free CPU instances (2 vCPU, 16 GB RAM), with paid upgrade options for GPU and other accelerated hardware including TPUs. [16]
Spaces supports multiple frameworks and SDK options:
| SDK | Description |
|---|---|
| Gradio | Python library for building interactive ML UIs with minimal code |
| Docker | Custom containers for arbitrary applications, APIs, and tools |
| Static | Simple HTML/CSS/JavaScript applications |
The community actively contributes to Spaces, which serves as both a portfolio for ML projects and a way for users to try out models directly in the browser without writing code.
Hugging Face maintains a large ecosystem of open-source libraries that cover the full ML workflow, from data loading and tokenization through model training, fine-tuning, and deployment.
The Transformers library is Hugging Face's flagship open-source project and is among the most popular ML libraries in the world. It provides thousands of pretrained models for tasks across NLP, computer vision, audio processing, and multimodal applications. The library provides a unified API for loading, configuring, training, and running inference on models from a wide range of architectures. [17]
Key features of the Transformers library include:
In 2025, Hugging Face released Transformers v5, a major update that adopted a PyTorch-first approach. The release sunset TensorFlow and Flax support in favor of deeper optimization for PyTorch, though the team has worked with partners in the JAX ecosystem to maintain compatibility through external libraries. Transformers v5 also introduced the "transformers serve" component for deploying models through an OpenAI-compatible API, streamlined inference with continuous batching and paged attention, and enhanced quantization as a first-class feature. [18][19]
The Datasets library provides efficient tools for loading, processing, and sharing AI datasets for NLP, computer vision, and audio tasks. Built on Apache Arrow, the library uses memory-mapped files for its local caching system, enabling users to work with datasets far larger than available RAM. [20]
Core features of the Datasets library include:
| Feature | Description |
|---|---|
| Zero-copy reads | Arrow format eliminates serialization overhead for fast data access |
| Memory efficiency | Memory-mapped storage allows loading datasets like English Wikipedia with only a few MB of RAM |
| Parallel processing | Configurable multi-process data preparation with automatic sharding |
| Streaming | Load and process datasets without downloading the full dataset to disk |
| Framework integration | Copy-free hand-offs to NumPy, Pandas, PyTorch, and TensorFlow |
| Processing operations | Built-in support for sampling, shuffling, filtering, mapping, and batching |
The Tokenizers library provides fast, production-grade tokenization implementations written in Rust with Python bindings. It can tokenize a gigabyte of text in under 20 seconds on a standard server CPU. [21]
The library supports multiple tokenization algorithms:
| Algorithm | Description |
|---|---|
| BPE (Byte-Pair Encoding) | Iteratively merges the most frequent character pairs |
| WordPiece | Used by models like BERT, splits words into subword units |
| Unigram | Probabilistic subword tokenization method |
| Character-level | Splits text into individual characters |
The library handles the full preprocessing pipeline including normalization, pre-tokenization, tokenization, and post-processing (truncation, padding, and adding special tokens). It also provides alignment tracking, making it possible to map any token back to its corresponding span in the original text. [21]
The Accelerate library simplifies distributed training for PyTorch models across different hardware configurations. It unifies common distributed training frameworks, including Fully Sharded Data Parallel (FSDP) and DeepSpeed, behind a single interface. [22]
Accelerate supports multiple parallelization strategies:
| Strategy | Description |
|---|---|
| Data Parallelism (DP) | Replicates the model across GPUs, distributes data batches, and synchronizes gradients |
| Fully Sharded Data Parallel (FSDP) | Shards model weights, gradients, and optimizer states across GPUs for memory efficiency |
| Tensor Parallelism (TP) | Distributes individual linear layer computations across devices |
| N-Dimensional Parallelism | Combines multiple parallelization strategies for maximum efficiency |
The library enables training on multi-GPU setups, TPUs, Apple Silicon, and other accelerated hardware, requiring minimal code changes to scale from a single GPU to multi-node clusters. [22]
The PEFT (Parameter-Efficient Fine-Tuning) library enables fine-tuning of large language models by training only a small number of additional parameters while keeping the base model frozen. This approach significantly reduces computational costs and memory requirements compared to full fine-tuning. [23]
PEFT supports several fine-tuning techniques:
| Method | Description |
|---|---|
| LoRA (Low-Rank Adaptation) | Inserts trainable low-rank matrices into model layers |
| QLoRA | Combines 4-bit quantization with LoRA for extreme memory efficiency |
| Prefix Tuning | Prepends trainable tokens to model inputs |
| P-Tuning | Learns continuous prompt embeddings |
| Prompt Tuning | Adds trainable soft prompts to the input |
| IA3 | Rescales model activations with learned vectors |
QLoRA is particularly notable because it allows models with up to 65 billion parameters to be fine-tuned on a single 48GB GPU while preserving the performance of traditional 16-bit fine-tuning. PEFT integrates with the Transformers, Diffusers, and Accelerate libraries. [23]
TRL (Transformer Reinforcement Learning) is a library for post-training foundation models using techniques like Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Group Relative Policy Optimization (GRPO). [24]
Key trainers and algorithms in TRL include:
| Trainer | Description |
|---|---|
| SFTTrainer | Supervised fine-tuning on instruction-following data |
| DPOTrainer | Direct Preference Optimization, used to train Llama 3 and many other models |
| GRPOTrainer | Group Relative Policy Optimization, used by DeepSeek R1, more memory-efficient than PPO |
| RewardTrainer | Trains reward models for RLHF pipelines |
| PPOTrainer | Proximal Policy Optimization for classic RLHF |
TRL leverages Accelerate for distributed training, integrates with PEFT for memory-efficient training via LoRA and QLoRA, and provides a command-line interface for fine-tuning without writing code. [24]
The Diffusers library provides state-of-the-art pretrained diffusion models for generating images, audio, video, and 3D structures. It offers ready-to-use inference pipelines, interchangeable noise schedulers for balancing speed and output quality, and tools for building custom diffusion pipelines. [25]
The library supports major generative models including Stable Diffusion, DALL-E variants, and others. Key features include text-to-image generation, image-to-image transformation, inpainting, negative prompts for controlling output, and CPU offloading for memory optimization. Diffusers integrates with PEFT for efficient model customization and follows a design philosophy prioritizing usability, simplicity, and customizability. [25]
Gradio is an open-source Python library that enables developers to build interactive web interfaces for ML models with just a few lines of code, requiring no JavaScript, CSS, or frontend experience. Originally created by Abubakar Abid in 2019 and acquired by Hugging Face in 2021, Gradio has become a cornerstone of the Hugging Face ecosystem. [7]
Gradio 5, released in October 2024, introduced AI-powered app creation, enhanced security features, and improved performance. The library supports a wide range of input and output types including text, images, audio, video, and 3D objects. Gradio applications can be shared via public URLs, embedded in webpages, and deployed to Hugging Face Spaces for permanent hosting. [26]
| Library | Purpose | Language | Key Feature |
|---|---|---|---|
| Transformers | Model loading, training, and inference | Python | Thousands of pretrained models |
| Datasets | Data loading and processing | Python | Arrow-backed memory-mapped storage |
| Tokenizers | Fast tokenization | Rust/Python | 1 GB of text in under 20 seconds |
| Accelerate | Distributed training | Python | Multi-GPU, TPU, FSDP, DeepSpeed |
| PEFT | Parameter-efficient fine-tuning | Python | LoRA, QLoRA, prefix tuning |
| TRL | Reinforcement learning from human feedback | Python | SFT, DPO, GRPO trainers |
| Diffusers | Diffusion model inference and training | Python | Stable Diffusion, image/video generation |
| Gradio | Interactive ML demos | Python | Browser-based model interfaces |
Hugging Face offers multiple inference options for running models in production. The serverless Inference API allows users to test models directly from the Hub without deploying any infrastructure. For production workloads, the company introduced Inference Providers in early 2025, a system that unifies over 15 inference partners (including Fal, Replicate, SambaNova, and Together AI) under a single, OpenAI-compatible endpoint. Users can switch between providers without changing their code. [27]
Inference Endpoints is Hugging Face's managed deployment service, designed for enterprises that need dedicated, scalable infrastructure for ML models. Users select a model from the Hub, choose a cloud provider and region, and specify security and scaling settings. The service supports any model from Transformers to Diffusers and provides auto-scaling, HIPAA compliance, GDPR compliance, and air-gapped environment options for regulated industries. [28]
Pricing starts at $0.032 per CPU core per hour and $0.50 per GPU per hour, billed per minute of actual usage. Enterprise plans offer dedicated support, 24/7 SLAs, uptime guarantees, and custom pricing based on volume commitments. [29]
Text Generation Inference (TGI) is Hugging Face's open-source toolkit for deploying and serving large language models. TGI powers Hugging Chat, the Inference API, and Inference Endpoints internally. It supports models including Llama, Falcon, StarCoder, BLOOM, GPT-NeoX, and T5. [30]
TGI has been influential in shaping the inference ecosystem. It initiated the practice of optimized inference engines relying on Transformers model architectures, an approach subsequently adopted by downstream engines like vLLM and SGLang, as well as local engines like llama.cpp and MLX. TGI now supports a multi-backend architecture that allows integration with different inference solutions, including vLLM as a backend. [30]
Hugging Face uses a freemium business model, with several tiers designed for different scales of use.
| Tier | Price | Key Features |
|---|---|---|
| Free | $0 | Public models and datasets, basic Spaces, community access |
| Pro | $9/month per user | Private models, advanced Spaces, early access to features |
| Team | $20/month per user | Centralized billing, team management, resource pools |
| Enterprise | $50+/month per user | Custom contracts, SLAs, dedicated support, elevated resource limits |
Enterprise customers receive managed billing with annual commitments, legal and compliance processes including custom contracts, personalized support with dedicated account management, and the highest storage, bandwidth, and API rate limits. [29]
Hugging Face also offers Expert Support as an add-on service for organizations that need assistance adopting the Hub, fine-tuning models, or deploying ML infrastructure. [29]
Hugging Face has co-led several large-scale open-science research collaborations that have produced significant AI models and tools.
The BigScience project was a year-long open research collaboration involving over 1,000 researchers from around the world. In 2022, the project released BLOOM (BigScience Large Open-science Open-access Multilingual Language Model), a 176-billion-parameter multilingual language model trained on 46 languages and 13 programming languages. BLOOM was the first open-source model to exceed the parameter count of GPT-3 and was released under the Responsible AI License (RAIL). [31]
BigCode is an open scientific collaboration, sometimes called the "spiritual successor" of BigScience, focused on responsible development of large language models for code generation. The project brought together over 1,200 members from institutions across 62 countries. In partnership with ServiceNow, the collaboration produced StarCoder, a 15.5-billion-parameter model trained on over 80 programming languages from The Stack dataset. StarCoder was released under an Open Responsible AI License (OpenRAIL). [32]
SmolLM is Hugging Face's family of small, efficient language models designed to run on resource-constrained devices. The original SmolLM series included models at 135M, 360M, and 1.7B parameters, pre-trained on SmolLM-Corpus, a curated collection of high-quality educational and synthetic data. [33]
SmolLM2 (1.7B parameters) showed significant advances in instruction following, knowledge, reasoning, and mathematics. It was trained on 11 trillion tokens from FineWeb-Edu, DCLM, The Stack, and new mathematics and coding datasets. SmolLM3, the latest generation, features a 3B-parameter model that outperforms Llama 3.2 3B and Qwen 2.5 3B while remaining competitive with larger 4B alternatives. SmolLM3 supports dual-mode reasoning (think/no_think) and multilingual output in six languages: English, French, Spanish, German, Italian, and Portuguese. [33]
Zephyr is a series of language models trained by Hugging Face's H4 (Helpful, Harmless, Honest, and Huggy) alignment team. Zephyr-7B-beta, the most notable release, is a fine-tuned version of Mistral 7B v0.1 trained on a mix of publicly available synthetic datasets using Direct Preference Optimization (DPO). The Zephyr models demonstrated that smaller, well-aligned models could achieve strong performance on conversational and instruction-following tasks. [34]
The Open LLM Leaderboard is an interactive benchmarking platform hosted on Hugging Face Spaces that allows the community to evaluate and compare open-source large language models. The leaderboard uses the Eleuther AI LM Evaluation Harness to run standardized benchmarks including IFEval, BBH, MATH, GPQA, MUSR, and MMLU-PRO. Around 300,000 community members use the leaderboard monthly through submissions and discussions. Version 2 of the leaderboard introduced updated benchmarks and improved reproducibility. [35]
Hugging Face has built one of the largest open-source AI communities in the world. With over 13 million users as of 2025, the platform has fostered a culture of open collaboration where users increasingly create derivative artifacts such as fine-tuned models, adapters, benchmarks, and applications, rather than simply consuming pre-trained systems. [14]
The company's community-driven governance model encourages transparency in model development and deployment. Model cards, dataset cards, and Spaces documentation provide structured information about capabilities, limitations, biases, and intended uses. This framework fosters accountability and helps ensure that AI technologies align with ethical standards. [2]
A curated feature called Tasks provides an organized view of models grouped by their intended purpose. For each task, the platform offers visual explanations with diagrams, videos, and links to interactive demos using the Inference API, along with descriptions of use cases and task variants. [2]
Hugging Face has been an active advocate for open-source AI in policy discussions. In March 2025, the company submitted a response to the White House Office of Science and Technology Policy's request for information on the AI Action Plan, arguing that open AI systems and open science are fundamental to making AI more performant, efficient, broadly adopted, and secure. [36]
The company has also advocated for open-source AI as a cornerstone of digital sovereignty, arguing that organizations and governments should have the ability to inspect, modify, and deploy AI systems independently rather than relying solely on proprietary platforms. [37]
Hugging Face occupies a unique position in the AI industry as a platform that bridges the gap between model discovery, development, and deployment.
| Platform | Primary Strength | Key Differentiator |
|---|---|---|
| Hugging Face | Model discovery and community | 2M+ models, largest open-source AI community |
| Replicate | One-click model deployment | Serverless, pay-per-use pricing |
| AWS SageMaker | Enterprise ML operations | Deep AWS integration, modular building blocks |
| Google Vertex AI | Managed ML on Google Cloud | BigQuery integration, TPU access |
| Azure AI | Enterprise AI on Microsoft Cloud | Integration with Azure services, OpenAI partnership |
Hugging Face's key competitive advantage is its network effect: models, datasets, and tools shared on the Hub attract users, who in turn contribute more resources, creating a self-reinforcing cycle. The company's strategy mirrors GitHub's early growth approach, building a massive free user base and then monetizing through enterprise features and managed services. [13]
Unlike cloud-native platforms like SageMaker or Vertex AI, Hugging Face is cloud-agnostic, allowing users to deploy models on any infrastructure. With the ggml.ai acquisition, the company now controls the full pipeline from model hosting (Hub), to model definition (Transformers), to local inference (llama.cpp), a combination that no other single organization offers. [9]
Several notable events have shaped Hugging Face's trajectory in 2025 and 2026:
Hugging Face has played a central role in the open-source AI movement. By providing free infrastructure for sharing models and datasets, the company has lowered barriers to entry for AI research and development. Its libraries have become standard tools in both academic and industry settings, and its Hub serves as the de facto repository for open-source AI models. [2][14]
The company's influence extends beyond technology. Its open research collaborations like BigScience and BigCode have demonstrated that distributed, community-driven research can produce models rivaling those from well-funded corporate labs. Its advocacy for open-source AI in policy discussions has helped shape the regulatory landscape around AI transparency and accessibility. [31][36]
With over 50,000 paying customers, $130 million in annual revenue, and a $4.5 billion valuation, Hugging Face has also demonstrated that building open-source AI infrastructure can be a viable business. The company's trajectory from a teenage chatbot startup to the central hub of the open-source AI ecosystem represents one of the more remarkable pivots in recent technology history. [8][12]