Parameters in AI: what they are, explained simply

In plain words

Think of a huge mixing desk in a recording studio, covered in sliders. Each slider changes the sound a little. Set them all just right and the song sounds perfect.

An AI model is like that, except the sliders are numbers and there can be billions of them. These numbers are called parameters, or weights. During training each one is nudged until the model's answers come out right. After that they are fixed and saved as the model file.

So when you read that a model has "70 billion parameters", it means 70 billion learned numbers. More parameters let a model hold more patterns, which is one reason bigger models can often do more. But size is not everything: data quality and training method matter just as much.

Everyday example

A family recipe as amounts

Your aunt's famous sambar comes down to amounts: how much tamarind, how much dal, how long on the flame, how many curry leaves. Change any one and the taste shifts. Over the years she tuned each amount until it was perfect. Parameters are those amounts. A model just has billions of them, all tuned together by training instead of by taste.

Where the example breaks: Each amount in a recipe has a clear meaning. A single parameter in a model means almost nothing on its own; what the model knows is spread across huge numbers of them at once.

How it actually works

In a neural network, the parameters are the weights on connections plus one bias for each neuron.

Weights and biases

Each connection between two neurons has a weight saying how strongly one affects the other. Each neuron also has a bias, a starting offset. Counting every weight and bias gives the parameter count.

Set by training, not by hand

No person types these numbers in. Gradient descent adjusts them all, step by step, to reduce the model's errors on its training data.

Size means memory

Each parameter is stored as a number, commonly 2 bytes. So a 70 billion parameter model needs roughly 140 gigabytes just to hold its weights, which is why quantization (storing fewer digits per number) matters for running models on smaller machines.

Not to be confused with hyperparameters

Hyperparameters are settings people choose before training, such as the learning rate or the number of layers. Parameters are what the model learns.

More is not always better

Scaling laws show bigger models improve predictably when given enough data and computing power. But a well-trained small model can beat a poorly trained large one, and mixture-of-experts models use only part of their parameters for each token.

AlexNet (2012)

About 60 million parameters

GPT-3 (2020)

175 billion parameters, ten times any earlier comparable language model

Llama 3.1 (2024)

Released in 8, 70 and 405 billion parameter sizes

As of October 2026

Many companies do not publish parameter counts for their largest models

Questions people ask

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

The B stands for billion parameters. A 7B model has about 7 billion learned numbers, a 70B model about 70 billion. It is a rough guide to how much the model can hold, and to how much memory and computing power it needs to run.

Are parameters and weights the same thing?

Nearly. Weights are the numbers on connections between neurons; biases are one extra number per neuron. Together they make up the parameters. In everyday talk, people use weights and parameters to mean the same thing.

Does a model with more parameters always perform better?

No. More parameters give more capacity, but results also depend on the amount and quality of training data and on the training method. Newer, smaller, well-trained models often beat older, larger ones.

How many parameters does ChatGPT have?

OpenAI has not said for the models behind ChatGPT, as of October 2026. It published 175 billion for GPT-3 in 2020, but its GPT-4 technical report gave no model size, citing competition and safety. Figures you see online for newer models are estimates.