What is a foundation model — and why an LLM is just one kind of it

You've already heard "LLM," "neural net," "model." Now "foundation model" keeps popping up too. It looks like just a fancy synonym. Here's the surprise: it isn't. An LLM is one kind of foundation model — a member, not the whole thing. In a couple of minutes you'll see why that's not word-splitting but an important distinction.

What it actually is

A foundation model is one huge neural network trained on everything at once, so it can later be adapted to hundreds of different tasks.

The key idea is two steps. First the model is trained a single time on a giant pile of data (the whole internet's text, millions of images, mountains of code). You get a "foundation" — a base understanding of the world. Then that foundation is adapted to a specific job: with a prompt, with fine-tuning, with document search. One foundation, many uses.

The term itself was coined at Stanford in 2021 — precisely to describe this new class of "trained once, used everywhere."

How it changed everything

To feel the shift, remember how it used to work.

Every task used to get its own model built from scratch: one for the spam filter, one for translation, one for spotting cats in photos. Each with its own data, its own training run, its own months of work.

Foundation models flipped that. You take one ready-made foundation — and it can already translate, summarize, answer questions, and write code. "Foundation" is literally the base: not a finished house, but a solid footing you can quickly build anything on. That's exactly why language models like GPT and Claude turned out so versatile — they're foundation models.

Why an LLM is just one kind

Now the key distinction. An LLM is a foundation model trained on text. But foundation models come trained on other material too.

• Trained on text → you get an LLM (writes and understands text).
• Trained on images and text together → a multimodal model (sees a photo and describes it).
• Trained on audio → a model for speech and music.

They're all foundation models. The LLM is just the most famous of the family, because text turned out to be the handiest medium. So "foundation model" is the family, and "LLM" is one member.

Why it matters to you

The practical takeaway: you almost never train a foundation model — you rent one. Training your own costs tens of millions of dollars and needs enormous hardware. But taking a ready one via an API and adapting it to your needs is within anyone's reach.

And when you hear "foundation model" in a product description, read it as "a general-purpose base, not a narrow single-task tool." Your job isn't to build the foundation — it's to build well on top of it: with a good prompt, your own data, the right settings.

Where you'll meet them

Every time you use ChatGPT, Claude, or Gemini — a foundation model is under the hood. An image generator like Stable Diffusion is a foundation model too, just for pictures. Almost any modern AI product is a thin layer on top of someone's big foundation.

Are a foundation model and an LLM the same thing?

No. An LLM is a foundation model trained on text. Every LLM is a foundation model, but not every foundation model is an LLM — there are also ones for images, audio, video. The LLM is the most popular member, which is why they get conflated.

Do I need to train my own foundation model?

Almost certainly not. That's a job for major labs: giant datasets, thousands of GPUs, millions of dollars. It's enough for you to take a ready model via an API and tune it to your task — with a prompt or your own documents. There's no reason to build the foundation from scratch.