temperature

advancedwidespread

Temperature is a setting that determines how predictable or how varied and surprising the language model's answer to the same question will be.

In plain terms

The language model assembles the answer word by word (more precisely, token by token), choosing from several possible continuations at each step. Temperature regulates how strictly it adheres to the most probable continuation versus allowing rarer, more unusual solutions.

Imagine a chef. At a low temperature, they always cook exactly by the recipe: reliable, but you always get the same dish. At a high temperature, they improvise: sometimes they find a great new flavor, sometimes the result is inedible.

That is why a low value is usually used for precise, verifiable tasks (data extraction, calculation, coding), and a higher one for brainstorming, naming and writing poetry. Temperature does not change what the model knows, only how it chooses from what it knows.

How it works

At each step, the model assigns a score to every possible next word fragment (token): the better it fits the text, the higher the score.
The scores are divided by the temperature value, and only then are they converted into probabilities. A value of 1 leaves the model's original probabilities unchanged.
If the value is small (close to 0), differences are amplified: the most likely token almost always wins, and the response becomes predictable. Division by exactly zero is impossible, so in practice 0 means the model always selects the most likely token without sampling.
If the value is high (above 1), the differences blur: rarer tokens also appear frequently, the text becomes more varied, but it is more prone to drift.
The model finally 'samples' a token from these adjusted probabilities, then repeats the entire process for the next step.

An everyday example

You ask the chatbot to name your new bakery, and each time you click the "Regenerate" button, you get different ideas. This is due to temperature: the system intentionally leaves some randomness in the response. If the same model is set up to extract data from invoices, it is run at a low temperature so that it produces much more consistent results.

Who is affected

everyoneThis is why you can get two different answers to the same question, and why it is worth verifying or re-asking important answers.

developerThe temperature of the model built into the application is usually set by you, and the reproducibility and testability of its behavior depends on this. However, in some newer models, the value is fixed or can only be adjusted within narrow limits. → Code

marketerFor creative texts and slogans, a higher value works better; for factual product descriptions, a lower value yields better results. → Marketing

entrepreneur, executiveWhen ordering a customer service or data processing AI solution, it is worth asking because an excessively high value leads to inconsistent responses.

A common misconception

MythIf I set the temperature to zero, the model does not err and always tells the truth.

In realityZero only means the model always chooses the most likely continuation, which can be confidently wrong. Fabricated claims (hallucinations) do not disappear at low temperatures, and the response is not always identical letter by letter.

Limits and risks

Low temperature does not guarantee accuracy: a consistent response can be consistently wrong.
An excessively high value can result in incoherent, meaningless text and more factual errors.
In most consumer chat applications, you cannot adjust this because the provider has pre-set it.
Values cannot be transferred from one model to another: the same number may represent a different scale and behavior with another provider.

Similar, but not the same

top-p sampling – Top-p completely excludes the least likely continuations. Temperature does not rule out any option; it only sharpens or blurs the differences between probabilities.
hallucination – A hallucination is the model's false, fabricated statement; temperature is merely a setting that may increase the likelihood of this, but is not the cause.
prompt – The prompt tells the model what you are asking, while the temperature determines how much freedom it has to choose among possible responses.

Related concepts

Broader

sampling, text generation, model parameter

Narrower

–

On the edge

greedy decoding, top-p sampling, top-k sampling, token, hallucination, large language model

A brief history

The name comes from statistical physics, where temperature determines how disorderedly particles are distributed among the possible states. It entered machine learning with the Boltzmann machines described in 1985, and the 2015 paper on knowledge distillation by Geoffrey Hinton and colleagues made it widely known.

Where things stand

As of late 2026, temperature is an adjustable parameter in the developer interfaces of most language models, but it typically stays hidden in consumer applications. For several newer step-by-step thinking (reasoning) models, the provider fixes or restricts its value, so it can be tuned little or not at all.

Status: 2026 November

In the news

Appears in 1 Agent Age stories; the curve shows weekly occurrence in the news magazine.

Schneider: warmer coolant could make AI data centers more water-hungrySep 21, 2026 · Ipar

Also known as

temperature · temperature · temperature parameter · temperature parameter · sampling temperature · sampling temperature · softmax temperature · LLM temperature

Sources

Hinton, Vinyals, Dean: Distilling the Knowledge in a Neural Network (2015)

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