What is a
model?
Strip away the mystique and an AI model is one thing: a very large box of numbers that has learned to guess the next piece of text. Get this one idea and everything else clicks into place.
When people say "AI," they almost always mean a model — something like GPT, Claude, or Llama. A model isn't a person, a database, or a search engine. It's a mathematical function: text goes in, a prediction comes out. The magic is entirely in how it makes that prediction.
A giant box of dials
Imagine a machine with billions of tiny dials. Each dial can be turned up or down, and the exact position of every dial determines how the machine responds to what you feed it. Those dials are called parameters (or weights), and a modern model has anywhere from a few billion to over a trillion of them.
Nobody sets those dials by hand — there are far too many. Instead, they're tuned automatically during training, by showing the model enormous amounts of text and nudging every dial, over and over, until its guesses get good. The finished model is those settings. "A 70‑billion‑parameter model" literally means 70 billion tuned dials frozen in place.
A model is just its weights. Copy the weights, you've copied the model. That's why a "model file" can be dozens or hundreds of gigabytes — it's the frozen positions of billions of dials.
Models don't read words — they read tokens
A model doesn't see letters or whole words. Text is first chopped into tokens — little chunks that are often a word, sometimes part of a word, sometimes just punctuation. "Understanding" might be one token; "oakweb" might split into "oak" + "web." A rough rule of thumb: one token ≈ ¾ of a word, so 1,000 tokens is about 750 words.
Everything a model does is measured in tokens: how much you send it, how much it writes back, and — as you'll see on the inference page — what it costs to run. When a provider quotes "$5 per million input tokens," this is the unit they mean.
It's autocomplete — taken to the extreme
Here's the whole trick: a language model does exactly one thing — it looks at the tokens so far and predicts the most likely next token. Then it adds that token to the text and predicts the next one. And again. And again. Stack that simple step billions of parameters deep, trained on a huge slice of human writing, and "guess the next word" turns into something that can write code, explain physics, and hold a conversation.
It is not looking anything up, and it has no live connection to the internet or a database unless a developer bolts one on. It is generating, token by token, from what its dials have learned. That's why it can be fluent and confident and still be wrong — a point worth remembering every time you use one.
How much it can "see" at once
A model can only consider so many tokens at a time — its context window. Think of it as the model's short-term memory or its desk: everything relevant to the task has to fit on the desk at once. Older models had room for a few thousand tokens; newer ones hold hundreds of thousands to over a million — entire books or codebases in a single go.
Two things follow from this. First, a model has no memory between separate conversations unless the app deliberately feeds the history back in — each request starts from a blank desk. Second, the context window is why techniques like retrieval (pulling in the right documents just before you ask) matter: you're choosing what goes on the desk.
The words that describe a model
The size
The count of tuned dials — 8B, 70B, 400B+. More parameters usually means more capable, but also heavier and slower to run.
The memory
How many tokens it can consider at once. Bigger context = it can read more of your material before answering.
The creativity dial
How much randomness when picking the next token. Low = focused and repeatable; high = varied and creative.
The model itself
The actual downloadable object (often GGUF/safetensors). Have the weights and the hardware, and you can run it yourself.
That's the whole foundation: billions of dials, tuned to predict the next token, working within a limited window. Next we'll watch those dials actually get tuned — that's training.
Common questions
What is an AI model in simple terms?
An AI model is a large mathematical function that has learned to predict the next piece of text. It's defined by billions of tuned numbers called parameters (or weights): feed it text, and it produces a prediction based on patterns it learned during training.
What are parameters and weights in an AI model?
Parameters (also called weights) are the adjustable values inside a model — think of them as billions of tiny dials. Their exact settings determine how the model responds. A "70-billion-parameter model" has 70 billion tuned dials. The finished model is essentially just its weights.
What is a token in AI?
A token is the small chunk of text a model actually reads — often a word, sometimes part of a word or punctuation. As a rough rule, one token is about ¾ of a word, so 1,000 tokens is roughly 750 words. AI usage and pricing are measured in tokens.
What is a context window?
The context window is the maximum number of tokens a model can consider at once — its short-term memory. Modern models range from a few thousand to over a million tokens. Anything outside the window isn't "seen" unless the app deliberately feeds it back in.