Definition
A foundation model is a single large AI model trained on broad data at scale that can then be adapted to perform many different tasks.
At a glance
- Trained once on broad data, then reused for many jobs instead of one model per task.
- Familiar examples: GPT-4, Claude, Gemini, and Llama[4].
- You adapt the general base with prompting or light fine-tuning on your own data.
- For a business: lower cost and faster results than building AI from scratch.
Why “foundation”
Stanford researchers coined the term in 2021[1]. One model acts as a shared base that many apps build on. Old AI needed a separate narrow model per task; one foundation model can power a chatbot, summarize contracts, and analyze reviews.
How a business uses one
You rent access from a provider rather than train your own[2]. Easiest path is prompting: describe the task in plain language. For deeper fit, fine-tune on a small set of your own examples, far cheaper than building from scratch[3].
What to weigh
They can give confident wrong answers, carry training-data bias, and send prompts to an outside vendor unless deployed privately. Decide which tasks need adapting, what data you will share, and whether prompting alone suffices before paying to fine-tune.
Bottom line
A foundation model is one general base you adapt rather than rebuild, so start with prompting and weigh cost, accuracy, and privacy.