Foundation model

A foundation model is a huge, general-purpose model trained once (at enormous cost) on broad data, then adapted for many different downstream jobs with relatively little extra work.

Think of it as the base layer:

• Train once, big: billions of parameters, trillions of tokens.
• Adapt many: fine-tune, prompt, or wrap it for chat, code, vision, etc.

That “train big, reuse everywhere” economics is why a few labs dominate — training a foundation model is one of the most expensive things a company can do.