Back to glossary Artificial Intelligence

Foundation Model

Large, pre-trained AI model serving as the foundation — customized via fine-tuning for specific applications.

What is a Foundation Model?

A foundation model is a large AI model pre-trained on massive datasets (text, images, code, audio) without specialization. Examples: GPT-4, Claude, Gemini, Llama. The foundation model is a "base" that's then adapted for specific applications.

From foundation to specialization

A bare foundation model is a generalist. Customization happens through: fine-tuning (retraining on domain data), RAG (providing context from knowledge bases), prompt engineering (system instructions defining role and constraints), and RLHF (learning from human feedback).

Open vs closed models

Foundation models come in open (Llama, Mistral — downloadable, runnable on your servers) and closed (GPT-4, Claude — accessible only via API). The choice affects cost, privacy, customization flexibility, and vendor lock-in risk.