Parameters (AI model)

When you read that a model has "7 billion parameters", that is the number of adjustable values inside its neural network. Together they store everything the model learned from its training data, though not as readable facts. Nobody can point to the one parameter that holds a particular fact.

Parameters are not the same as settings you control. Options such as temperature are sometimes loosely called parameters, but they only shape how the model picks its output; they do not change what it learned. More parameters usually means more capability and higher running costs, but not always better answers.

How it works

Each parameter is a number that controls how strongly one unit in the network influences another. Training starts with random values and nudges every one of them, over and over, until the model's predictions match the examples well. Once training ends, the parameters are fixed and saved as the model file.

Running a model means doing calculations with all those parameters for every token it produces, which is why large models need powerful chips and cost more per answer. Parameter counts are useful for comparing models within one family, where the same company used similar data and methods. Across families the comparison is weaker, because a smaller model trained on better data, or for longer, can outperform a larger one. Some models also activate only part of their parameters for each token, so the headline number overstates how much computing each answer actually uses.

Parameter count also decides memory. Each parameter is stored as a number taking a set number of bytes, so a model's file size and the memory needed to run it grow with its parameter count. A technique called quantisation stores each parameter with fewer bits, shrinking the model so it fits on cheaper hardware, at some cost in quality. This is how some open-weight models can run on a well-equipped laptop.

Example

A web agency compares two models for a support chatbot. The larger one writes slightly better replies but costs more and responds more slowly. A small language model with far fewer parameters handles the shop's narrow set of questions just as well, so the agency picks the smaller one.

To decide, the agency writes fifty real customer questions, runs both models on them and scores the answers for accuracy and tone. Measured that way, the size difference matters less than expected, and the result gives a reason for the choice that the team can repeat when newer models arrive. They also keep the fifty questions as a fixed test set, so any future model can be compared on the same ground.

Why it matters for your business

Parameter count helps explain price and speed differences between models, and whether a model can run on a laptop or needs a data centre. For most business tasks, the right question is not which model is biggest but which is the smallest one that does your job reliably. Smaller models also cost less to run at volume, respond faster for visitors and can sometimes run on your own server, which helps when data must stay in-house. The trade-off is weaker handling of unusual or complex requests, so test with the awkward questions your customers really ask, not only the easy ones. Ask any supplier which model size they chose and what they tested it against.

Common mistakes

• Choosing a model by parameter count alone rather than by testing it on your real tasks.
• Confusing a model's parameters with the adjustable settings in an AI tool's interface.
• Assuming companies always publish parameter counts; many keep them private.

Questions people ask

Does a model with more parameters give better answers?

Usually it can handle harder and broader tasks, but not always better for your task. Training quality, data and tuning matter as much as size. Smaller models trained well often match larger ones on narrow jobs, while being faster and cheaper to run.

What does 7B or 70B mean in a model name?

It is the approximate number of parameters, in billions. A 7B model has about seven billion parameters and a 70B model about seventy billion. The number gives a rough sense of the hardware needed to run the model, especially for open-weight models you host yourself.

Can I change a model's parameters?

Only by training it further, a process called fine-tuning, which needs either the model's weights or a provider's fine-tuning service, plus a good set of examples. Most businesses instead shape a model's behaviour with prompts, examples and retrieved documents, which leaves the parameters untouched and is far cheaper.