Glossary
Foundation model
A foundation model is a large general-purpose model trained once at great expense on broad material, then adapted by many organisations for many different jobs instead of each of them training something from scratch.
In plain terms
It is the base layer everything else is built on. Making an AI system used to mean training something from nothing for each job: one thing for sorting email, another for reading invoices, each needing its own gathered examples and its own expense. Now one very large general model is trained at enormous cost, and everybody else starts from that and adapts. The expensive step moved to a handful of organisations, and the second step became cheap enough for a small company to take.
Why it matters
It explains the shape of the market you are buying in. A small number of organisations can afford to train these, so most of the products you will evaluate are a thin layer over one of a few shared bases. That concentration is worth understanding before you sign anything, because it decides what happens to your tool when a provider changes its terms, retires a version or raises its prices. It is also why a capability seems to appear in a dozen unrelated products in the same month: they did not each invent it.
How it works
The training run is deliberately not aimed at a task. The material is broad rather than curated for a job, and what comes out is something generally capable instead of something good at one thing. That generality is the whole point and the reason the same base can sit under a coding tool, a customer support assistant and a document summariser without being retrained for any of them.
Adaptation happens afterwards and costs a tiny fraction. Standing instructions, a few examples inside the request, retrieval of your own documents, or a light additional training pass are all ways of pointing a general base at a particular job. None of them needs anything like the computing power the original run consumed, which is what puts them within reach of ordinary organisations.
The economics concentrate hard. A training run at this scale needs computing resources available to relatively few organisations, so the number of genuinely distinct bases is small compared with the number of products built on them. When you compare five vendors in a category, it is quite normal to find they resolve to two or three underlying bases and differ in everything above that line.
They come in families, in sizes, and with different ranges of input. Some handle text only, some also take images or audio, and providers publish some for anyone to download while keeping others available only as a service. Which of those applies to a given base decides whether running it on your own hardware is an option at all.
One base, and everything built on it
What changed about where the money goes
Seen in the wild
Use one general assistant for a contract summary, a spreadsheet formula and a marketing email in a single sitting, none of which it was specifically trained to do.
ClaudeSend a request through a routing interface offering a long list of models, and notice how short the list of organisations behind them actually is.
OpenRouterDownload a published general-purpose model to your own machine and point it at whatever you like, which is the adaptation half of the pattern done by you rather than by a vendor.
Ollama
Common misconceptions
People assume
Any company selling AI trained its own model.
In fact
Very few did, and building on somebody else's base is the normal, sensible thing to do. But our AI and our product built on a provider's base are different claims with different risks attached, and only one of them survives that provider changing its pricing. Ask which, during evaluation rather than after.
People assume
Foundation model is another way of saying language model.
In fact
The term describes a role rather than a medium. Bases handling images, audio or several kinds of input at once are foundation models too. Language ones are simply the famous ones at the moment, which makes the two words look interchangeable when they are not.
People assume
It already knows my industry.
In fact
It knows whatever appeared in broad public material, which for a specialised field is often shallow, dated or absent. Anything current, internal or genuinely specialist has to be supplied at the time you ask. Depth in a niche is nearly always the adaptation layer doing the work rather than the base.
Telling them apart
Foundation model vs Large language model
Foundation model
Describes the role: a general base that many downstream things are built on. Says nothing about what kind of input it takes.
Describes the medium: it works on text. Says nothing about whether anyone else builds on it.
Most language models people talk about are also foundation models, which is why the words get swapped. A small text model trained for one narrow job is a language model and no kind of foundation.
Foundation model vs Fine-tuning
Foundation model
The shared starting point. Identical for everyone who builds on it, and far beyond most organisations to produce.
One of the ways of making that starting point yours, by continuing its training on your own examples.
The base is what everybody has; the adaptation is what differs. When a vendor's advantage is real it is nearly always in the second, and asking which one they mean by their technology is a fair question.
Questions
- Who actually trains these?
- A small number of well-resourced organisations, because a training run at this scale needs computing power and specialist staff beyond most companies. That is the defining economic fact about the term: the expensive step happens once, in few places, and everyone else starts downstream of it.
- How do I find out which base our vendor uses?
- Ask directly, and ask two follow-ups: whether they can change it without telling you, and whether you can be pinned to a version. A vendor unwilling to answer has made your capability and cost assumptions unstable, which matters more than the specific answer they would have given.
- Does building on somebody else's base make a product less good?
- Not at all. Nearly every product worth buying does it, and the useful work is usually in the layer above: what the tool retrieves, what it instructs, what it checks. The thing to watch is not the dependency but whether the vendor has acknowledged it and planned for the provider changing course.
- What happens if the provider retires the model we depend on?
- The product moves to a successor, and behaviour changes with it, sometimes noticeably. This is a real continuity risk rather than a theoretical one. Ask a vendor how much notice they give, whether their prompts and adaptations are portable to another base, and whether they have ever had to move before.
Key takeaways
- A foundation model is a general base trained once at great expense and adapted by many organisations afterwards.
- The term names a role in the market, not a medium, so image and multi-input bases count too.
- Bases are few and products are many, so competing vendors often resolve to the same two or three.
- Adaptation is the cheap half, which is where a vendor's real advantage usually sits.
- Depending on someone else's base is normal; not knowing which one, or whether it can change, is the risk.
Tools that use this
- Claude
One general base doing unrelated jobs in a single sitting.
- OpenRouter
A long catalogue of models resolving to a short list of organisations.
- Ollama
Published bases you can download and adapt yourself.
Last checked July 2026