Model, LLM, foundation model

Three words, three claims

Model, foundation model and LLM get used as if they were the same word with different amounts of jargon attached. They are not synonyms - they nest, each one narrower than the last, and each makes a different claim about the thing it describes.

So every LLM is a foundation model, and every foundation model is a model. The reverse fails in both directions: a decision tree that predicts churn is a model and nothing more, and an image generator is a foundation model that is not an LLM.

The term worth being slightly careful with is foundation model. It is a claim about how something was built - pretrained broadly, adapted afterwards - and it gets applied loosely in marketing to anything large. When a vendor calls something a foundation model, the question worth asking is what it was pretrained on and what it is meant to be adapted to.

Reasoning models

This is the label that causes the most confusion, because it sounds like a different kind of system and is not. A reasoning model is an LLM that has been trained to generate a long internal working-out before its visible answer - and that is given the time and tokens to do it.

The mechanism from What is an LLM? is unchanged: still next-token prediction, still no lookup. What changed is that the model produces a great deal of intermediate text first, and that intermediate text conditions the final answer. Working through a problem step by step genuinely does improve results on tasks with a chain of dependent steps - maths, multi-file code changes, planning.

Two consequences that matter in practice. It costs more and takes longer, because those thinking tokens are generated and billed like any others. And the gain is uneven: strong on problems with real intermediate structure, close to nothing on recall or simple formatting. Most current frontier models expose it as a dial rather than a separate product, which is the more honest framing - reasoning is a mode, not a category.

Multimodal models

Multimodal means the model handles more than one kind of data - most commonly text plus images, increasingly audio and video.

The part that trips people up is that it is two claims, not one, and vendors rarely separate them: what the model can read, and what it can produce. A model that accepts a screenshot and replies in text is multimodal on input only, which describes most of the assistants people use daily. A model that generates images is multimodal on output. These are different capabilities with different prices, and "multimodal" on a spec sheet tells you neither on its own.

Strictly, a model that reads images is no longer purely a language model, which is why "LLM" is drifting toward being an umbrella term for the whole class rather than a precise description. That drift is now standard usage, and correcting people on it is a losing battle - but it is worth knowing that the L has quietly stopped being accurate.

What the words do not tell you

None of these labels is a quality signal, and three things people read into them are not there:

• Size. "Large" in LLM is relative and decades out of date as a threshold. Model sizes now span several orders of magnitude and vendors mostly stopped publishing them.
• Capability. Nothing in these words distinguishes a frontier model from something small and cheap. Both are LLMs.
• Openness. Orthogonal to all of it. Open-weight and closed models sit in exactly the same categories.

What actually separates one model from another is benchmark performance on tasks resembling yours, context window, price, latency and availability. The category words tell you what kind of thing you are looking at, which is a genuinely useful thing to know and the limit of what they offer.

So which word do you use

A short rule for each:

• Model - when the specific kind does not matter, or when the audience is non-technical. Almost always safe.
• LLM - when the point is that it works on text and predicts tokens. The right word for explaining behaviour: context limits, hallucination, cost per token.
• Foundation model - when the point is that one broadly pretrained artifact underpins many applications. Useful in architecture and procurement conversations, mostly noise elsewhere.
• Reasoning model - when the distinction affects a decision, because it changes latency and cost. Better still, name the setting rather than the category.
• Multimodal - only with the direction attached. "Accepts images" and "generates images" are the useful statements; the bare word is not.

The general habit worth having: these are category labels, not specifications. When a sentence needs to carry a real claim - it reads PDFs, it costs this much, it holds this much context - say that instead. The category word is for orienting a listener, and it stops being useful the moment a decision depends on it.