Back to glossary

Term

Hugging Face

Hugging Face is the central platform of the open-source AI community — a hub for models, datasets, and demos, and the maintainer of widely used libraries like Transformers, Diffusers, and PEFT.

Hugging Face — explained

Hugging Face started as an NLP library (Transformers) and is now the de-facto standard for open AI models. The Hub hosts over a million models (Llama, Mistral, Qwen, Gemma, DeepSeek, Stable Diffusion, Whisper, and many more), hundreds of thousands of datasets, and interactive Spaces (demos built on Gradio or Streamlit). Around it sits an ecosystem of open-source libraries: Transformers (model access), Diffusers (image models), PEFT (parameter-efficient fine-tuning), TRL (RLHF/DPO), Datasets (efficient loading), and TGI (inference server).

Example / Practical context

Loading a Llama 3.1 8B model locally means pulling it from meta-llama/Llama-3.1-8B-Instruct on the Hub — either as PyTorch weights (Hugging Face format) or, for the llama.cpp world, as a GGUF variant of the same file from a community repo. Spaces like chat.lmsys.org (LMSYS Arena) or Whisper demos run on Hugging Face. Commercially, there are Inference Endpoints (hosted models), AutoTrain (no-code fine-tuning), and enterprise plans.

The Hugging Face Hub is the platform, Transformers is the Python library — often used interchangeably, but technically different. Unlike OpenAI or Anthropic, which offer proprietary models behind API endpoints, Hugging Face is primarily about distribution: models from many providers are shared, mirrored, and versioned. For local inference, tools like Ollama or LM Studio compete — but Hugging Face usually supplies the source files.

Entdecke mehr

Themenuebersicht