What Is a Language Model

You have probably used a chat app like ChatGPT, Claude, or Gemini. You type something. An answer appears.

Behind that chat box sits a program. That program is a language model.

People also call it an LLM — short for large language model. In this guide, model, language model, and LLM mean the same thing.

The chat box is only the face. The model is the engine that writes the words.

One job

A language model does not dig through a library of saved answers like a search page.

It does one simple job: look at the text so far, and guess what should come next.

Then it guesses again. And again. Those guesses become the full reply you read.

Watch one short reply get built

Suppose the text so far is:

The cat sat on the

The model does not “look up” a finished sentence. It picks a likely next piece, for example:

The cat sat on the mat

Then it looks at that longer text and picks again:

The cat sat on the mat .

Then it may stop because the sentence feels finished.

So a reply is not fetched whole. It is grown, piece by piece. Those small pieces are what we will call tokens in a moment.

Why the same question can sound different

Ask twice:

Say hello in one short sentence.

One run might answer:

Hello — nice to meet you.

Another run might answer:

Hi there!

Both are fine greetings. The model is choosing likely next pieces each time, not opening a fixed page titled “hello.” Small differences in those choices change the wording.

What tokens are

Those next pieces have a name: tokens.

The model does not think in whole pages. It works in small tokens — pieces of text.

Rough feel (not a hard law):

So the guessing loop is really:

1. Read the tokens so far
2. Pick a good next token
3. Add it to the text
4. Repeat until the answer is done

Example idea with a tiny question:

2 + 2 =

The model may continue with tokens that become:

2 + 2 = 4

You still read normal words. Under the hood, the model is choosing the next small piece, over and over.

Why this matters later: longer text means more tokens. More tokens usually means more work for the server — and more cost when you pay for use.

Where the model lives

The model usually runs on a company’s server — a computer on the internet that answers requests.

Your own program (or the chat app) sends your text to that server. The server runs the model. A reply comes back over the internet.

You do not need to build or train the model yourself to use it. Someone else already trained it and put it on a server you can call.

  Your program                      Server
  ┌──────────────┐   your text     ┌──────────────┐
  │  send        │ ──────────────▶ │  model runs  │
  │              │ ◀────────────── │              │
  └──────────────┘   the reply     └──────────────┘

Concrete picture:

1. Your program sends: What is 2 + 2?
2. The server runs the language model on that text
3. Your program receives a reply such as: 2 + 2 is 4.

Same pattern whether a chat app does it for you, or your own program does it later.

What it cannot do by itself

Because the model only sees the text you send it, it cannot reach outside that text on its own.

Example: you ask, “What is in my notes.txt?”

If you did not paste the file contents into the request, the model has no way to open your disk. It may guess. Guessing is not reading your file.

So remember the whole picture in one line:

A language model continues text, piece by piece (in tokens), on a server you call over the internet — and it only knows what you show it.

Cast so far: