Temperature
A sampling setting that controls how random a language model's output is — low values make responses focused and repeatable, high values make them more varied and creative.
Temperature scales the probability distribution a language model samples from. At each step the model produces a score for every token in its vocabulary; temperature divides those scores before they are turned into probabilities.
- Temperature 0 (or near zero): the model almost always picks the highest-probability token. Output is deterministic and flat, which is useful for extraction, classification, and code.
- Temperature ~0.7: a common default that balances coherence and variety for chat and writing.
- Temperature 1.0+: flattens the distribution, making unlikely tokens more probable. Useful for brainstorming, risky for facts.
Related controls often appear beside it: top-p (nucleus sampling) keeps only the most probable tokens whose cumulative probability passes a threshold, and top-k restricts sampling to the k best candidates.
Practical notes: temperature does not make a model smarter — it only changes which plausible continuation is chosen. For reasoning tasks, some systems deliberately use low temperature to keep chains of thought on track, while creative work benefits from higher settings. The setting also interacts with repetition: too high, and coherence breaks down; too low, and answers become repetitive.