Hugging Face is a platform where the machine learning community shares models, datasets and applications. Its center is the Hugging Face Hub, which works like a code-hosting site for AI: each model or dataset is a repository (a versioned folder of files, here mostly model weights, the numbers a trained model consists of), with a public page.
Anyone can host unlimited public models, datasets and apps. Spaces run small web demos of models on free CPU or on paid GPUs billed by the hour. Inference Endpoints deploy a chosen model from the Hub on dedicated servers that scale with traffic, and Inference Providers give one API to 45,000+ models from several providers. Hugging Face also maintains open-source libraries, among them Transformers for using pretrained models in PyTorch, Diffusers for image models, Datasets, Tokenizers, PEFT and TRL for fine-tuning (further training of an existing model on new data).
Public hosting is free. A PRO account, Team and Enterprise plans, and GPU computing are paid; compute is billed per hour of use.