Hugging Face

Hugging Face started as a chatbot company in 2016 and pivoted to become the infrastructure layer for open-source AI. Its core product is the Hub: a repository where any organization or individual can upload model weights, datasets, and demo applications. When Meta releases Llama 4, Google releases Gemma, or Mistral releases Large 3, the weights typically appear first or simultaneously on Hugging Face. The Hub has more than 1 million models as of 2025 and is the de facto distribution point for open-weight AI.

The Transformers library (also maintained by Hugging Face) is the most widely used Python library for working with large language models: loading weights, running inference, fine-tuning, and integrating models into applications. Alongside Transformers, Hugging Face maintains PEFT (parameter-efficient fine-tuning, including LoRA), Datasets (standardized dataset access), Evaluate (benchmarking), and TRL (reinforcement learning from human feedback). Spaces is Hugging Face's hosting product for interactive demos, built with Gradio or Streamlit.

For builders, Hugging Face is unavoidable if you work with open-weight models. It's where you find models, compare benchmarks (via the Open LLM Leaderboard), run quick experiments in Spaces, and download weights for local deployment with tools like Ollama or vLLM. The Inference API lets you call hosted models via API without running your own GPU infrastructure, which is useful for prototyping. Hugging Face is also a company: it has raised significant venture funding and runs paid enterprise services (including private model hosting and inference endpoints), but its public-facing hub remains free to use.