AI temperature and Top P explained for roleplay and stories
What the two dials actually change in the next sentence, which values to use for play, and the three other settings that matter more than people think.
Temperature decides how far the AI strays from the most expected next word, and Top P decides how many words it is allowed to choose from at all. For roleplay and stories, as of November 2026, a temperature between 0.8 and 1.2 with Top P left at 1 covers almost everything. Most "the AI got weird" problems come from a different setting, covered below.
What does temperature do in AI roleplay?
Every time the model writes a word, it has a ranked list of candidates with a probability on each. Temperature reshapes that list. At low values the top candidate wins almost every time, so the model writes the safe, predictable sentence. At high values the lower candidates get real chances, which reads as inventive until it stops making sense.
The default is 1.0. In play, 0.8 to 1.2 covers almost every world. Below 0.7 the prose turns flat and repeats its own phrasing, and the AI stops surprising you. Above 1.3 names change mid-scene and rules get forgotten, because the model picked an unlikely word and then committed to it.
Raise it for slice-of-life, romance and comedy, where surprise is the point. Lower it for mysteries, strategy games and anything where the AI has to keep track of numbers. Yumina calls this dial creativity.
The creator of a world may pin a temperature for it, because the world's rules were tuned at that value. If the dial is greyed out, that is why.
What does Top P do?
Top P trims the candidate list before temperature touches it. At 1, every candidate is in play. At 0.9, the model keeps only the likeliest words whose probabilities add up to 90 percent and throws the rest away, which is the long tail of odd choices. Temperature then picks from what is left.
Leave it at 1 unless the model keeps producing a specific kind of nonsense, a made-up word or a wrong language, at a temperature you otherwise like. Then lower Top P to 0.9 or 0.95 and keep the temperature. Lowering both at once makes the model dull.
Which settings matter more than temperature?
Most complaints blamed on temperature come from these four.
Story memory
How much of your conversation the AI reads each turn. When it is too small, the AI forgets scene one by scene fifteen, and that forgetting looks like randomness. Can AI remember a long roleplay? covers it.
Response length
A ceiling on one reply, 12,000 tokens by default on Yumina. Lower it if the AI writes a page when you wanted a paragraph. The stronger fix is the opening, because the AI copies the length of what the creator wrote first.
Reasoning effort
On models that think before they write, this is how long they think. Low is right for play, because it keeps replies fast and the prose natural. High helps when the model keeps breaking the rules of the world, at the cost of slower turns.
Frequency and presence penalty
Two dials that push the model away from words and ideas it has already used. A small frequency penalty, 0.2 to 0.4, stops the "she smiled softly" loop that long chats fall into. Presence penalty nudges it toward new topics. Both are off by default and most worlds never need them.
Does the AI model matter more than the settings?
Yes. A stronger model at temperature 1.0 keeps rules that a smaller model loses at 0.7. On Yumina, the model pill in the chat box switches the AI for the next reply, and Settings > AI Config sets your defaults. Choosing an AI model has the tiers and the costs.
What temperature and Top P should you start with?
| What you want | Temperature | Top P |
|---|---|---|
| A rules-heavy game, numbers tracked | 0.7 to 0.9 | 1 |
| A story that follows the world closely | 0.9 to 1.0 | 1 |
| Romance, comedy, slice of life | 1.0 to 1.2 | 1 |
| The model keeps inventing words | keep yours | 0.9 |
Change one dial at a time and play three turns before judging it. The same value reads differently on different models.
FAQ
What is the best temperature for AI roleplay?
Between 0.8 and 1.2 for most stories. Use 0.7 to 0.9 for a rules-heavy game that tracks numbers, and 1.0 to 1.2 for romance, comedy and slice of life.
Should I change temperature or Top P?
Change temperature first and leave Top P at 1. Lower Top P to 0.9 or 0.95 only when the model keeps inventing words or switching language at a temperature you otherwise like.
Why does the AI keep repeating itself?
Below a temperature of 0.7 the prose repeats its own phrasing. In long chats, a frequency penalty of 0.2 to 0.4 stops the repeated lines.
Why does the AI change names or forget rules mid-scene?
Above a temperature of 1.3 the model picks an unlikely word and commits to it. If it forgets scene one by scene fifteen, the cause is story memory, and temperature will not fix it.
Sources
Checked 14 November 2026 against Choosing an AI model, which documents the defaults quoted here. The value ranges are what creators on Yumina report; they are starting points.