parameter

basicwidespread

A parameter is a number inside an AI model, set by the model during training; billions of them together carry its knowledge.

In plain terms

Imagine a huge mixing console with billions of tiny knobs. Each knob is a parameter, meaning a single number. How these knobs are currently set determines what the model answers to your question.

Humans do not turn these knobs. During training, the model sees countless examples and adjusts the knobs slightly after every mistake until its answers improve. The 'knowledge' of the finished model is therefore not stored in sentences or rules, but in these numerous adjusted numbers.

The number of parameters is often referred to as the model size: a "7 billion parameter" model is considered small, while one with hundreds of billions is large. More parameters generally mean greater knowledge, but also higher computational requirements and costs.

How it works

A model is an artificial neural network: small computing units arranged in layers and interconnected. Each connection has a number, a weight, which indicates how much that connection matters. Most units also have a number called a bias. The weights and biases together form the parameters.
At the start of training, parameters are random numbers, so the model still gives incoherent answers.
During training, the model receives examples, guesses the answer, and the program measures how far off it was.
A mathematical procedure calculates which parameter to shift, in which direction, and by how much to reduce the error. This step is repeated countless times during training, even hundreds of thousands or millions of times for large models.
At the end of training, the parameters are fixed. When you ask the model a question, these no longer change: the model only calculates with them to produce the answer.

An everyday example

You read a news item about a new '8 billion parameter' language model that runs on a laptop. This number indicates the model is relatively small: faster and cheaper, but expected to know less than its cloud-based counterparts with hundreds of billions of parameters.

Who is affected

everyoneIn the news, parameter count is the most common metric for model size, so it is worth knowing what it means and what it does not.

entrepreneur, executiveThe parameter count strongly influences how much it costs to run a model and whether it fits on your own machines or is only accessible via the cloud.

developerWhen selecting a model, you can estimate memory requirements, speed, and whether it is strong enough to back an agent from the parameter count. → Code

parent, student, teacherIt helps to understand that AI does not read answers out of an encyclopedia but generates them from learned numbers, so it can also make mistakes. → Education & Family

A common misconception

MythThe more parameters a model has, the smarter it is.

In realityThe parameter count only shows the size. Quality depends at least as much on the training data, the training method, and the model architecture, so a newer, smaller model is often better than an older, larger one.

Limits and risks

The parameter count alone does not reveal what a model is good for: test results and your own trials are worth checking.
Parameters cannot be read like text: even researchers only partially understand which number is responsible for what, making it difficult to explain individual answers.
Large companies often do not disclose their models' parameter counts, so some figures circulating in the press are estimates.
The model's knowledge is 'frozen' into the parameters at the end of training, so it cannot know about later events without separate search or updates.

Similar, but not the same

Hyperparameter – The developer sets the hyperparameter before training (for example, the learning rate), while the model itself learns the parameter.
token – A token is a piece of text read or written by the model, while a parameter is the model's internal, learned number.
setting (e.g., temperature) – Settings provided during use are sometimes called parameters, but these only tune the response style and do not change the knowledge the model has learned.

Related concepts

Broader

artificial intelligence, machine learning, Neural network

Narrower

weight, shift

On the edge

training, Hyperparameter, large language model, fine-tuning, token

A brief history

The word came from mathematics and statistics into machine learning. The idea of adjustable weights during learning appeared in Frank Rosenblatt's 1958 perceptron. The parameter count became a recurring metric in the news in 2020 with the launch of the 175 billion parameter GPT-3.

Where things stand

At the end of 2026, parameter count is the standard metric of model size. Models that run on phones work with at most a few billion parameters, while the largest may have several hundred billion or even more than a trillion. Many large models use a so-called mixture of experts architecture. Such a model uses only a small share of its parameters at each step of a response, so the total and the actually used parameter counts can differ significantly. Attention is increasingly shifting toward actual performance and cost, and large companies often do not disclose the number at all.

Status: 2026 November

In the news

Appears in 5 Agent Age stories; the curve shows weekly occurrence in the news magazine.

DeepSeek's agent platform reaches 1.54 million daily users as Chinese open models spreadNov 2, 2026 · Agent Age
Mistral Large 4 is the best open model outside China, with weights released publiclyOct 29, 2026 · Agent Age
Meta releases the open-weight Muse Spark Open, with 816 billion parameters, and Muse WonderOct 12, 2026 · Agent Age
Mistral unveils 1-trillion-parameter Large 4, plans open-weights releaseOct 6, 2026 · Agent Age
Chinese Qwen search tool removed from US Federal Register site; Alibaba releases new Qwen image modelSep 17, 2026 · Agent Age

Also known as

parameter · parameters · parameter count · model parameter · weights · parameter · model parameters · weights

Sources

Brown et al. (2020): Language Models are Few-Shot Learners (the GPT-3 paper)

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.