Foundation model
A large general-purpose model trained once at great expense, then used as the base for many different products.
What it tells you about a developer
A candidate who worked on a foundation model was inside one of a few dozen companies, and the work is research-flavored. A candidate who built on one did product engineering, which is a far larger pool and a far more common req.
A foundation model is trained on a broad sweep of data and is not built for any one task. GPT, Claude, Gemini, and Llama are foundation models. Training one costs millions of dollars and takes months, so only a handful of companies do it. Everyone else builds on top, either through an API or by downloading an open-weight model.