If you have used ChatGPT or Claude, you have used a model without anyone calling it that. It is worth two minutes to understand what one actually is, because almost everything people get wrong about AI — and most of what it ends up costing them — follows from getting this wrong.
A tool, not an assistant
A model is a program that was shown an enormous number of examples until it got good at one narrow job. That is the part people get wrong. A model is not an all-purpose AI. It is closer to a single tool on a workbench: a saw does not do what a drill does, and neither of them sands anything.
The chat assistants everyone knows — ChatGPT, Claude, Gemini, Copilot — are language models: trained on text, good with words. They are one kind among many. The model that lifts a product off its background has never read a sentence in its life. It was trained on photographs, and it does exactly one thing.
You can teach one your own job
This is the part most people never find out, and for a business it is the part that matters. A model is not only something you pick off a shelf and hope fits. Almost any of them can be taught your specific work.
You do not build one from scratch. That genuinely is a research-lab exercise — millions of dollars, a team of specialists, months of compute — and anyone implying otherwise is selling something. What you do instead is take one of the free public models and show it a few hundred examples of your job done the way you want it done. It learns the pattern, and afterwards it does that job without being told how each time.
In practice that looks like:
- A model that writes quotes in your format, with your product names and your terms, because it has read four hundred of your old ones.
- An image model that has learned your product photography look from twenty photographs, so next season's shots match the catalogue instead of nearly matching it.
- A page reader taught your supplier's particular invoice layout, so the right numbers land in the right columns.
- A transcriber that has learned your trade's vocabulary and stops turning it into nonsense.
Two things before that sounds too good. It wants a few hundred consistent examples — below about two hundred a model tends to memorize your examples rather than learn the pattern behind them, and inconsistent examples teach it to be inconsistent. And the useful part: what you produce is a file you keep and can run anywhere. There is more on the practicals, including what it costs, in teaching a model your material.
Nine kinds you will actually meet
There are more than nine, but these are the families that cover nearly all real work. Only the first is the kind you have already used, and the last three are the ones a business tends to hit first.
The distinction between the second and third is the one that costs people the most. Making an image and changing an image are different families of model, trained on different things for different purposes. Ask a generator to clean up your photograph and it will hand you a new photograph that resembles it.
Not all nine are equally worth teaching your own material. Words, making images, reading pages and speech are the four where a few hundred of your own examples change the result noticeably, and they are where most business fine-tuning happens. Changing images and video are usually used as they come — the general models are already good at those jobs and teaching them is expensive for little gain.
And one worth separating out, because it is the most common expensive mistake here. When somebody says they want “a model that knows our documents”, they almost never need training at all. They need finding meaning: their material indexed so it can be searched by what it means, with a general model reading the results. That takes hours rather than a training run, costs less, and updates the moment a document changes. Train a model to change how it works; index your documents so it knows what you have.
They are public, and there are millions of them
Most models are published openly by research teams and companies — Alibaba, Tencent, Meta, universities, independent developers — and anyone can download them for free. They are kept in public libraries the same way code is kept on GitHub.
- Hugging Face is the main one by a wide margin. Millions of models covering text, images, video and audio, each with a page explaining what it does.
- Civitai focuses on image generation, and especially on the style add-ons that make a model produce a particular look.
- GitHub holds much of the surrounding code — the tools that run a model, chain several together, and handle the files.
Free and public sounds like the end of the story. It is the start of it. Nobody can evaluate millions of options, and the names give you nothing to go on. Worse, most of what is in there is not worth your time: by Hugging Face's own accounting, about half the models on the platform have been downloaded fewer than two hundred times ever, while a couple of hundred models account for roughly half of all downloads.
So it is not really millions of choices. It is a few dozen good ones, buried in millions.
Some models need more machine than others
This is the part that surprises people, and it is worth getting right because it is not about speed.
A model has to fit in a machine's memory before it can run at all. If it does not fit, it does not run slowly — it does not run. Speed only decides how long something takes once it already fits.
The model that removes a background from a photograph is small enough for an ordinary laptop. The one that generates eight seconds of video is not, and no amount of waiting will change that. It is closer to whether a piece of furniture goes through the door than to how fast you carry it.
That is the whole of the constraint. Everything else about machines is detail you can safely ignore until you have a reason not to.
How people actually run them
There are several routes, and what they cost differs sharply depending on how much work you have: paying per image or per word through somebody else's service, buying a card of your own, or renting one by the hour. Which is cheapest flips depending on volume, and it flips harder than most people expect.
That comparison deserves its own page rather than a paragraph here. What matters for now is that the choice exists and it is not obvious.
What this means if you just want something done
You do not have to choose a model. Describing what you want in plain English is enough, and the matching happens for you.
What is worth carrying away is the shape of the thing: several narrow tools rather than one clever assistant, most of them not the chat kind, chained together in an order that matters. That is why two jobs which sound identical can need completely different models, and why "can AI do this" almost always has a more interesting answer than yes or no.
If you would rather see it happen than read about it, how it works walks through the ninety seconds step by step.
Questions people ask
What is an AI model, in one sentence?
A program that was shown an enormous number of examples until it got good at one narrow job. Not an all-purpose assistant - closer to a single tool on a workbench, where a saw does not do what a drill does and neither of them sands anything.
Is an AI model the same as a large language model?
No. A large language model is one kind of model, the kind trained on text — ChatGPT, Claude, Gemini and Copilot are all this kind. There are also models for making images, editing photographs, reading scanned pages, transcribing speech and generating video. Assuming all AI is chat-shaped is the most expensive misunderstanding in this field.
Where do AI models come from?
Most are published openly by research teams and companies - Alibaba, Tencent, Meta, universities, independent developers - and anyone can download them for free. They live in public libraries, the largest being Hugging Face, the same way code lives on GitHub.
Can I get a model that does my company’s specific job?
Yes, and not by building one from scratch — that is a research-lab exercise costing millions. You take a free public model and show it a few hundred examples of your job done your way, and it learns the pattern. The result is a file you keep and can run anywhere.
How many examples do I need to teach a model my work?
A few hundred, and consistency matters more than volume. Below roughly two hundred a model tends to memorize your examples instead of learning the pattern behind them, and examples that contradict each other teach it to be inconsistent. Two hundred good ones beat a thousand careless ones.
If models are free, why would I pay for anything?
Free to download is not free to run, and picking the right one out of millions is its own job. A model needs a machine with enough memory to hold it, and most real work needs several models chained in the right order rather than one.
Do I need to choose a model myself?
No. Describing what you want in plain English is enough, and the matching happens for you. Knowing roughly how the categories work is still useful, because it tells you when something is possible at all.