Foundation model
A foundation model is a large-scale AI model trained on broad, diverse data and then adapted to a wide range of specific tasks — the backbone of modern generative AI.
Foundation models (term coined by Stanford in 2021) are the backbone of the modern AI revolution. They're trained once on petabytes of data (text, code, image) and then specialised via fine-tuning, RLHF or prompting. GPT-4, Claude 3.7, Gemini 1.5 and Llama 3 are all foundation models. Characteristics: large (billions of parameters), expensive (millions to hundreds of millions of dollars to train), broadly applicable.
Example
A fintech builds a compliance chatbot on top of Claude as foundation model. Fine-tuning on internal procedures and legal texts yields a model that specifically answers AML, PSD2 and MiCA questions — without training from scratch.
Frequently asked questions
Foundation model, LLM or frontier model?
Foundation model is the broadest term (any modality). LLM is a subset (text only). Frontier model = the biggest and most capable at a given moment.
Open-source or closed-source foundation models?
Closed (GPT, Claude, Gemini): best quality, API-dependent, data leaves you. Open (Llama, Mistral, DeepSeek): free, self-hostable, quality closing in. Mixed use is normal.
Related terms
Further reading
- → Our service: GEO