Skip to content

Glossary

Temperature

Temperature is a setting controlling how predictable a model's output is, from taking the safest continuation every time to letting less likely ones through, which reads as variety or as unreliability depending on the job.

In plain terms

Every time the model produces the next piece of text it has a ranked list of candidates to choose from. Turned down, it takes the top one every time, so asking twice gives you nearly the same answer twice. Turned up, it sometimes takes the second or the fifth, and the answers spread out. What it never does is change how right the model is. It changes how adventurous the picking is, not the quality of what there is to pick from.

01

Why it matters

Set wrongly it produces a problem that looks like a broken model and is a two-minute fix. An extraction step running at a lively setting will read the same invoice twice and give you two different totals, which sends people off investigating the wrong thing entirely. Turned down for writing, the same tool produces the same three tired openings for a month. Because it is often buried in a settings panel or chosen for you by a product, it is one of the more common causes of a tool behaving oddly for reasons nobody thinks to check.

02

How it works

At each step the model has candidate continuations with scores attached, and this setting reshapes those scores before one is chosen. Turned down, the distribution is squashed towards the leading candidate, so it wins almost always. Turned up, the gap between candidates narrows and less likely ones start getting picked. Nothing about the underlying ranking changes; only how decisively the leader is favoured.

At or near zero you get about as close to repeatable as this technology offers. The same request produces very nearly the same answer each time, which is what you want for anything a colleague might have to reproduce. It is not a cast-iron guarantee, because other things in a hosted service can introduce small variations, so treat it as reliable rather than certain.

It has no bearing on whether the answer is right. A confidently wrong reply at zero comes back identically wrong on every run, which is genuinely useful for diagnosis and is not an improvement in accuracy. If a model is missing the information it needs, no setting on this dial will supply it.

A second setting, usually presented as top-p or nucleus sampling, does an overlapping job by limiting how far down the candidate list the choice may reach. Adjusting both at the same time is a reliable route to confusing results, and the ordinary advice is to leave one at its default and move the other.

Where a product does not expose it, someone has chosen on your behalf, usually somewhere in the middle for general use. That single hidden decision explains a good deal about whether a tool feels consistent, and it is a fair question to ask a vendor when a workflow needs the same input to produce the same output.

Where four ordinary tasks want to sit

Where four ordinary tasks want to sitThe setting is a dial rather than a switch, and the right position follows from what failure looks like for the task. Reading totals off an invoice fails if two runs disagree, so it belongs hard against the left. Sorting a message into a queue is the same case with slightly more tolerance. Drafting a reply wants a little room, because a tool producing byte-identical letters is unsettling to receive twice. Generating twenty subject lines fails in the opposite direction entirely: twenty near-identical suggestions are worth no more than one. Notice that nothing on this line is about which end produces better work. It is about which kind of wrongness would hurt, and that question has a different answer for each of these four.MUST REPEAT EXACTLYSAMENESS IS THE FAILUREPulling totalsout of aninvoicetwo runs mustagreeSorting asupportmessage into aqueuesame message,same labelDrafting areply to acustomersome room,still onmessageTwenty subjectlines tochoose fromtwentynear-identicalones areuseless
The setting is a dial rather than a switch, and the right position follows from what failure looks like for the task. Reading totals off an invoice fails if two runs disagree, so it belongs hard against the left. Sorting a message into a queue is the same case with slightly more tolerance. Drafting a reply wants a little room, because a tool producing byte-identical letters is unsettling to receive twice. Generating twenty subject lines fails in the opposite direction entirely: twenty near-identical suggestions are worth no more than one. Notice that nothing on this line is about which end produces better work. It is about which kind of wrongness would hurt, and that question has a different answer for each of these four.
03

Seen in the wild

  • Send one request twice at two different settings through the same routing interface and see how far apart the two answers land.

    OpenRouter

  • Turn it down on an automation step pulling fields out of documents, so the same invoice cannot produce two different totals on two runs.

    n8n

  • Run a model on your own machine with the control in front of you and watch one sentence loosen as you raise it.

    LM Studio
04

Common misconceptions

People assume

Turning it down makes the model more accurate.

In fact

It makes the model more repeatable, which is a different property and often more useful. The most likely answer is not the correct answer; it is the one the model rates highest. Turning it down makes the same answer very likely to come back twice, and says nothing at all about whether that answer is right.

People assume

Turning it up makes the model more creative.

In fact

It makes the output more varied, and variety is an ingredient of creative work rather than the thing itself. Pushed far enough the text stops cohering altogether. If you want unexpected ideas, a better instruction produces more of them than a higher setting does.

People assume

High temperature is what causes made-up answers.

In fact

It can widen the spread of a wrong answer, and it is not the cause. Fabrication comes from the model having no reliable basis for a claim and producing something anyway, which happens at zero just as readily. The fix is supplying the material, not adjusting the dial.

05

Questions

What should I set it to?
Near zero for anything that must be reproducible: extraction, classification, routing, structured output, anything a colleague might rerun. Moderate for drafting and conversation, where identical phrasing every time reads oddly. High only for deliberate idea generation, and expect quality to fall as you push it.
Why does the same request give me different answers?
Most likely because the setting is above zero, which is the default in most conversational products. Turning it down is the first thing to check when variation is the problem. If answers still move at zero, the cause is elsewhere, in a changed system prompt, a different model, or retrieved material that was not the same.
My tool does not expose it. Does that matter?
It matters when you need the same input to produce the same output, and rarely otherwise. Where a product hides it, ask what it is set to and whether it can be fixed for your account. A vendor unable to answer is telling you that reproducibility is not something their product offers.
Should I change it or change my prompt first?
The prompt, nearly always. This dial governs how much the model wanders among plausible outputs; the instruction governs what plausible means. Most problems people try to solve here, including dull writing and inconsistent formats, are better solved by saying what you want more precisely.
06

Key takeaways

  • It controls how adventurously the model picks among candidates, not how good the candidates are.
  • Near zero gives you repeatability, which is not the same thing as accuracy.
  • A wrong answer at zero comes back identically wrong, which is useful for diagnosis.
  • Adjusting it alongside a second sampling setting is a reliable way to confuse yourself.
  • Where it is not exposed, a product has chosen for you, which explains a lot about consistency.
07

Related terms

08

Tools that use this

  • OpenRouter

    Two settings on the same request, side by side.

  • n8n

    Unattended extraction steps, where repeatability is the whole requirement.

  • LM Studio

    The control is in front of you while a model runs locally.

Last checked July 2026

All glossary terms