What do 'parameters' and model size mean?

A parameter is one of the tunable numbers inside a neural network, the weights that decide how strongly signals pass from one neuron to the next. Training sets their values from data. Model size usually means how many parameters a model has, often millions or billions. More parameters give more capacity to learn patterns, though bigger is not always better.

If you have read our guide on what a neural network is, you have already met these numbers under another name. Let us look at them properly.

What is a parameter?

Picture the neural network from that earlier guide: layers of neurons, joined by connections, each connection carrying a strength called a weight. A parameter is one of those weights, a single tunable number that shapes how a signal flows through the network. A large model holds a great many of them.

These numbers are not written by a person. They start out more or less random, and training adjusts each one a hair’s breadth at a time, guided by example after example. By the end, the settled values are what the model has learned. Everything the model can do lives in these parameters, and nowhere else.

So when you hear that a model has “billions of parameters”, that is a count of these tunable numbers. There is no hidden library of facts alongside them. The parameters are the model.

What does “model size” mean?

When people describe a model by its size, they almost always mean its parameter count. A model might be described as having millions of parameters, or billions, and that figure is the headline number you see in write-ups and comparisons.

It is a measure of capacity, a rough sense of how much room the model has to store patterns. Think of it a little like the number of dials on a very large machine. More dials mean more settings the machine can hold, and so more subtle behaviour it can learn to produce.

That analogy is imperfect, and it is worth saying so. The count tells you how many dials there are, not how well they were set. A machine with more dials is not automatically better tuned than one with fewer.

Does more parameters mean a better model?

Not on its own. More parameters do give a model more capacity to capture patterns, and at the frontier of what AI can do, scale has clearly mattered. So the instinct that bigger is more capable is not baseless. It is only incomplete.

There are real costs on the other side. A bigger model takes more computing power and more time to train, and every answer runs through all of those parameters, so it costs more to run and can be slower to respond. For a straightforward task, a huge model is overkill, like hiring a moving lorry to carry a single box.

And a smaller model often wins. One trained on better, cleaner data, or tuned for a specific job, can beat a much larger general model at that job. Size sits alongside two things that matter at least as much: the quality of the training data, and the technique used to train and refine the model. A large language model is shaped by all three, not by parameter count alone.

How should I use the parameter count?

Treat it as a rough hint, not a verdict. A bigger number suggests more capacity and, alongside it, more cost and more waiting. A smaller number suggests something lighter and quicker that may well be enough. Neither figure tells you which model is best for what you are doing.

Here is a practical way to hold it in mind:

• Read it as capacity and cost, together. More parameters mean more potential and more expense. The two travel as a pair.
• Do not read it as a score. A larger model is not automatically better than a smaller one at your task.
• Weigh it against data and tuning. A model trained on better data, or tuned for your kind of work, can beat a bigger rival.
• Match the size to the job. Simple, well-defined tasks rarely need the largest model available.
• Judge by results, not the spec sheet. The model that gives you the best answers is the right one, whatever its size.

Try it yourself

Next time you see two models compared, look past the parameter counts. Take a task you actually care about, a kind of email you write often, or a type of question you tend to ask, and run it through a smaller, faster model as well as a large one. You can find ready-made prompts to try in our prompt library.

Compare the answers side by side. You may find the smaller model is more than enough for everyday work, and quicker to boot. That hands-on comparison teaches you far more than any headline number.

Common mistakes

A common first instinct is to reach for the biggest model for everything, on the assumption that more parameters must mean better answers. You will often get a better experience by starting with a smaller, faster model and moving up only if the results fall short.

Another easy slip is treating the parameter count as the whole story. It is one factor among several, and the quality of a model’s training data and tuning can matter more. When you see AI terms used loosely, our glossary is there to help you keep them straight.

Next steps

You now know what parameters are and what a model’s size does and does not tell you. To see these numbers doing their work, read our guide on how an LLM actually works, which follows the signal from your words through the parameters to the reply on your screen.