AI 101

Temperature

Also known as: temp

Inference

Temperature is a number (usually 0โ€“1, sometimes a bit more) that controls how much the model “rolls the dice” when picking each next token.

Every step, the model has a ranked list of candidate words. Temperature decides how much weight the less-likely candidates get:

TemperatureBehaviorGood for
0.0โ€“0.3Nearly always picks the most likely word. Deterministic, safe.Code, JSON, factual answers, agents
0.5โ€“0.7Balanced. Varied but on-topic.General chat, summaries, most writing
0.8โ€“1.2+Frequently takes the fun riskier word.Brainstorming, stories, taglines

So the model doesn’t have a “creativity mode” you turn on โ€” you’re just widening the set of acceptable next words.

Rule of thumb: when an AI writes code or data, use a low temperature. When it writes copy or ideas, nudge it up. Most chat apps default to ~0.7, which is fine for both.