Glossary
Model
A model is the trained file at the heart of an AI product, holding everything it learnt from its training material, with the app or chat window you actually use wrapped around it.
In plain terms
The model is the engine and the product is the car. The engine decides how much power there is; the seats, the steering, the dashboard and the satnav decide what driving it is like. Two cars with the same engine can be nothing alike to sit in, and a good engine in a badly built car is still a bad car. Notice that this sense of the word is not the everyday one: a financial model is a simplification of something real, whereas this is a specific file you can copy.
Why it matters
You are usually choosing two things at once and only being shown one. The model decides the ceiling on capability, and the product around it decides almost everything about whether that capability reaches your work. Vendors sometimes change the model underneath a product without changing the product's name, which is why a tool can quietly get better or worse between one month and the next. Knowing the two are separable lets you ask which one is responsible when something disappoints, and that question has different answers with very different remedies.
How it works
A model is a very large collection of numbers, learnt during training, together with the code needed to run them. Those numbers are its entire contents: there is no separate store of facts, no index, no database. Given an input it produces a likely output, and that single narrow operation is all it does. Everything else you experience is built around it by somebody else.
It is fixed once training finishes. The file that answers your question this afternoon is byte for byte the file that answered somebody else's this morning, and nothing you typed changed it. Where a product appears to remember you, the product has stored something and is putting it back in front of the model next time, which is a feature of the software rather than a change to the model.
Models come in families and in sizes within a family. A larger one in the same family is generally more capable, and reliably slower and dearer to run, so products often route routine requests to a smaller sibling and reserve the large one for harder work. That routing is usually invisible, and it is one of the reasons the same product can feel inconsistent from one request to the next.
Most of what you experience is the product rather than the model: the interface, the standing instructions it sends on your behalf, the documents it retrieves for you, the tools it can reach, the checks it applies to the answer. This is why asking which model a product uses is a real question with a real answer, and still only half the story.
Which half is responsible when something disappoints
Seen in the wild
Send the same request to several different models through one routing interface and compare the replies, which isolates the model as the only thing that changed.
OpenRouterDownload a published model to your own machine and watch how much disk it takes, which makes the point that this is a concrete artefact rather than an abstraction.
OllamaLoad two models into the same local chat application and put an identical question to each, which separates what the model contributes from what the app around it does.
LM Studio
Common misconceptions
People assume
The model learns from my conversations.
In fact
It does not. Training finished before you met it, and the file is unchanged by anything you type. Products that seem to remember you are storing something separately and supplying it again next time. Whether your conversations are later used to train a future model is a different question, answered by the provider's terms rather than by the software.
People assume
It means a simplified representation, as in a financial model.
In fact
Not in this field. Here the word names a specific artefact: a file of learnt numbers that can be copied, stored, versioned and run. The everyday sense of the word describes an abstraction; this sense describes a thing on a disk, and confusing the two makes most explanations harder than they need to be.
People assume
A bigger model is always the better choice.
In fact
Bigger is slower and dearer, and on a great many everyday tasks a smaller one in the same family produces work you could not tell apart. The honest approach is to try the cheaper option on your own material first and move up only where the difference shows, rather than buying the largest available by default.
Telling them apart
Model vs AI
Model
A specific artefact: one trained file that can be named, copied, run and compared with another.
A category label covering many methods and products, with no particular artefact behind it.
You can ask which model, and get an answer. Asking which AI gets you a product name at best. When a claim resists being pinned to a named artefact, it is usually a claim about the category.
Model vs Large language model
Model
The general word. Covers models handling images, speech, numerical prediction and anything else trained the same way.
One kind of model, specialised in text, which happens to be the kind behind most of what is currently sold as AI.
Every large language model is a model; most models are not language models. The words are used interchangeably at the moment because the text ones are the famous ones, and that will not stay true.
Questions
- Where does the model actually live?
- Usually on the provider's servers, with your request travelling there and the answer travelling back. Published models can also be downloaded and run on hardware you control, in which case the file sits on your own disk. Which of the two applies is the single biggest factor in where your data goes.
- Can we choose which model a product uses?
- Sometimes. Assistants aimed at individuals often let you pick, and business tools built on a provider's interface frequently do not, choosing on your behalf and changing that choice without telling you. It is worth asking during evaluation, because a product that will not say has made the answer to your capability questions unstable.
- Does the model know about my company?
- Only whatever appeared in its training material, which for most organisations is very little and may be out of date. Anything current or internal has to be supplied with the question at the time you ask. A product that appears to know your business is fetching that material, not recalling it.
- How often do models change?
- Frequently, and often silently. A product can move to a newer model without renaming anything, which is why a tool that suited a task in one quarter may behave differently in the next. If a workflow depends on consistent behaviour, ask the vendor whether you can pin a version and how much notice you get before a change.
Key takeaways
- A model is a concrete file of learnt numbers, not a simplification in the everyday sense of the word.
- It is fixed after training and unchanged by your conversations; apparent memory is the product storing things.
- Larger models in a family cost more and run slower, and are often indistinguishable on routine work.
- Most of what you experience is the product built around the model, not the model itself.
- Products can change the model underneath without changing their name, which is worth asking about.
Tools that use this
- OpenRouter
One interface across many models, so the model is the only variable.
- Ollama
Downloads models to your own machine, where the file is visibly a file.
- LM Studio
Swap models under an unchanged interface and see what the app was contributing.
Last checked July 2026