Hugging Face

Hugging Face, Inc. is an American machine learning company that develops computation tools for building applications with machine learning. Its core product is the Hugging Face Hub, a collaboration platform on which users host, discover and share models, datasets and applications ("Spaces"); the site advertises more than 2 million models, over 1 million applications and more than 500,000 datasets, spanning text, image, video, audio and 3D modalities. The models directory reported 3,011,222 models listed. The company also maintains a widely used open-source stack, including Transformers (164,293 GitHub stars as listed on its site), Diffusers, Tokenizers, Datasets, Accelerate, PEFT, TRL, Safetensors, Transformers.js, Text Generation Inference, smolagents and the Hub Python library.

Beyond hosting, Hugging Face develops its own models. It leads the SmolLM family of small language models: SmolLM (2024, 135M/360M/1.7B parameters), SmolLM2 (135M/360M/1.7B, aimed at on-device use) and SmolLM3 (2025, a 3-billion-parameter multilingual reasoning model with long-context support covering six languages). It also publishes SmolVLM, compact vision-language models released with weights, training recipes and datasets under Apache License 2.0. Historically, the company convened the BigScience Research Workshop (launched April 2021), which produced BLOOM, a 176-billion-parameter multilingual large language model, in 2022. Other model work has included StarCoder and SafeCoder with ServiceNow and an open chatbot released with the nonprofit LAION.

Commercially, Hugging Face sells Team and Enterprise plans starting at $20 per user per month with single sign-on, regions, priority support, audit logs, resource groups and a private datasets viewer; managed compute through Inference Endpoints and GPU-backed Spaces starting at $0.60/hour; and an Inference Providers layer giving unified API access to 45,000+ models from third-party providers with no service fees. Earlier paid functionality described by TechCrunch included AutoTrain (automated model training), the Inference API and Infinity (production inference acceleration), alongside SaaS and on-premises versions of the hub.