GuidesUnderstanding AI
Temperature and top‑p: the settings of an AI model
What temperature controls
For each token, the model has a list of possible continuations with their probability. Temperature sharpens or flattens that list. Low, it almost always takes the favourite. High, it dares the next ones.
The other settings
- Top-pKeeps only the continuations that add up to a given share of probability.
- Max lengthCuts the answer beyond a number of tokens. A cost ceiling.
- Stop sequencesA text that, if it appears, stops generation.
Which value for which use
| Use | Temperature | Why |
|---|---|---|
| Extract fields, classify | Low | The same document must give the same result |
| Customer service answers | Low to medium | Consistent, yet natural |
| Write a text | Medium | Varied without drifting |
| Suggest ideas | High | Diversity is the goal |
What temperature does not do
It does not make an answer more accurate. A mistake can be repeated at low temperature. Reliability comes from sources and checks: see hallucinations.
Frequently asked questions
- Temperature 0: is the answer always identical?
- Almost, with no absolute guarantee. For a strictly reproducible result, validate the output in code rather than rely on temperature.
- Should temperature and top-p be tuned together?
- No. Providers advise changing one or the other, not both.
- Do all models accept these settings?
- No. Some recent so-called reasoning models ignore or fix these parameters. Check the documentation of the model you pick.