Hugging Face

Hugging Face is the open-source AI platform that hosts models, datasets, and applications — serving as the GitHub of machine learning. Founded in 2016 by Clément Delangue, Julien Chaumond, and Thomas Wolf, Hugging Face has become the central hub for the open-source AI ecosystem. Its Hub documentation, as of October 2026, describes the platform as hosting over 2 million models, 1.5 million datasets, and 1.5 million AI apps (Spaces).

The Model Hub

Hugging Face's Model Hub is where the AI community shares, discovers, and deploys language models, image generators, speech models, and every other category of machine learning model. Models from Meta (LLaMA), Mistral, Google, and thousands of independent researchers are hosted, versioned, and made accessible through standardized APIs. This creates a composable AI ecosystem: developers can swap models, combine them in pipelines, and fine-tune them for specific tasks.

Hub Scale and Open-Model Distribution

The public model listing showed 3,128,462 models on 6 October 2026, and the public dataset listing about 1.08 million; the documentation's headline figures and the live listings do not match exactly, so any single number should be read as approximate and dated. What matters more than the count is the Hub's function as the default release channel for open-weight models. Labs publish weights as Git-based repositories with version history, a model card describing intended use and limitations, and a machine-readable licence tag, and the card has become the de facto primary source for a model's architecture, context length and vendor-reported benchmarks. The Qwen organisation alone listed 468 model repositories when checked, and the DeepSeek-V4 technical report points readers to a Hugging Face collection for its checkpoints, as the Qwen repository does for its models.

Distribution at this scale has side effects worth knowing. Each popular base model accumulates a tree of derivatives — the Qwen3.6-35B-A3B page listed several hundred fine-tunes and adapters within six months of release — and download counters make adoption visible in a way closed APIs do not. Large files are stored on a chunk-deduplicating backend the documentation calls Xet, and a serverless Inference Providers API gives programmatic access to hosted models. The Hub also carries its own risks: licences vary per repository, benchmark numbers on cards are self-reported, and the documentation lists malware scanning among its security features because uploaded artefacts cannot be assumed safe.

Transformers Library

Hugging Face's Transformers library is the most widely used open-source ML library in the world, providing a unified API for working with transformer-based models across tasks (text, vision, audio, multimodal). The library's design philosophy — simple, consistent interfaces for diverse model architectures — has been instrumental in democratizing access to state-of-the-art AI.

PEFT and TRL: The Post-Training Libraries

Two further libraries cover adapting models after pre-training. PEFT (Parameter-Efficient Fine-Tuning) is described in its documentation as a library for adapting large pretrained models "without fine-tuning all of a model's parameters", training only a small number of extra parameters; it is the reference implementation for methods such as LoRA and integrates with Transformers, Diffusers and Accelerate (see parameter-efficient fine-tuning). TRL (Transformers Reinforcement Learning) is a full-stack post-training library whose documented trainers include supervised fine-tuning (SFTTrainer), GRPO (GRPOTrainer), direct preference optimization (DPOTrainer), reward modeling and distillation. Its GRPO trainer can hand generation to vLLM, either as a separate server or co-located on the training GPUs. TRL reached version 1 in March 2026, according to the project's blog index. Together with the Hub, these libraries mean that the common path from an open base model to a fine-tuned, published derivative runs through Hugging Face tooling end to end.

Spaces and Community

Hugging Face Spaces allows anyone to deploy ML applications (using Gradio or Streamlit) with a single click. This has created a vibrant community of AI demos, interactive research papers, and lightweight applications. The platform embodies the creator economy for AI: researchers and developers share work, build on each other's contributions, and collaborate in the open.

Open-Source vs. Closed AI

Hugging Face is the institutional champion of open-source AI. As debate intensifies between open and closed model approaches, Hugging Face provides the infrastructure that makes open-source AI viable — hosting, serving, and community tools that allow open models to compete with the closed APIs of OpenAI and Anthropic.