Hugging Face

Hugging Face is a platform for hosting, finding and deploying open-weight models, with routed LLM inference, dedicated endpoints and per-seat team plans.

Hugging Face gives machine learning teams one place to find an open model, try it and put it into production. Hugging Face Inc. runs the platform together with Hugging Face SAS, its French company. Developers come to it when they want to build on models they can inspect and download rather than on one closed model. The Hub is the core: repositories for models and datasets, each model documented on a model card that can carry evaluation results. Spaces host Gradio and Docker apps for demos, with ZeroGPU sharing GPU time among them and Spaces Dev Mode on PRO. Inference Providers routes chat completion, function calling, structured output and feature extraction requests to partner companies through one account. Inference Endpoints deploys a chosen model on dedicated, autoscaling hardware in AWS, Azure or Google Cloud, using serving engines such as vLLM, TGI and TEI. For training, the open-source TRL library covers supervised fine-tuning, DPO and GRPO. The platform suits ML engineers who ship open-weight models, researchers who publish their work, and startups that need a model behind an API without building the serving stack. Larger organizations get single sign-on with SAML or OIDC and Storage Regions to decide where repositories live. Individuals start on a no-cost account and can move to a personal PRO subscription, while companies pay per seat on Team or Enterprise. Compute and storage are billed by usage, by the hour or by the terabyte, and Enterprise contracts run through sales. A typical week on the platform mixes several jobs. A team compares candidate models in the browser, then calls the winner through Inference Providers on Groq, Cerebras or Together AI. Once usage climbs, the team deploys that model on Inference Endpoints, which do not store request payloads. Researchers fine-tune a checkpoint with TRL, push the weights back with a model card, and publish a Space so others can try it. It fits teams that want open models, hosted demos and production inference under the same login.