What are parameters?

Picture a tech headline announcing a new artificial intelligence (AI) model with 70 billion parameters, as if you'd know what to do with that number. Most people don't, and that's fine. It's a count of something real. And if you've heard how a neural network learns, you've already met the thing being counted.

A neural network is made of simple units joined by connections, and each connection has a strength: a number that makes a signal count for more or for less. Each of those numbers is a parameter. So are a few extra numbers, called biases, that nudge each unit's output up or down. A model with 70 billion parameters has 70 billion of these adjustable numbers inside it. That's all the headline is counting. If you read them aloud, one a second, without stopping, you'd be at it for more than two thousand years.¹²

Nobody sets them by hand. No person could. Training does it. The model makes a guess, the guess is measured against the right answer, and every parameter is moved a tiny amount toward a better guess. That happens over and over, across an enormous amount of data. When training ends, the parameters are fixed, and in a real sense they are the model. Everything it learned is stored in those numbers.²

The count does tell you two useful things.

• Room: more parameters give a model more room to store patterns, which can let it handle more kinds of tasks.
• Cost: every parameter has to be stored, and used in the math each time the model answers, so a bigger model needs more memory and more computing power to run.

That's why the largest models run in data centers, on rows of specialized chips, and why your laptop can't hold them.¹

What the count can't tell you is how good the model is. In 2022, researchers at DeepMind trained a model called Chinchilla with 70 billion parameters. It beat their own earlier model, Gopher, which had 280 billion, on a wide range of tests. The difference was training: with the same computing budget, Chinchilla learned from about four times as much data. How much data a model sees, how good that data is, and how it's tuned afterward can matter as much as its size.³

Size is a trade, and the right size depends on the job. Small has advantages of its own. A model small enough to fit on a phone can answer without sending your request to a data center, so it can work offline and keep more of your information on the device. Apple, for example, said in 2025 that the model built into its newer devices has about 3 billion parameters, small next to the largest models, and designed to run efficiently on the device itself. So when the next headline leads with a number, you can ask a better question. Big enough for what, and running where?⁴