Temperature

A setting that controls randomness in a model's output — low for consistent, deterministic answers, higher for varied, creative ones.

Temperature scales how much randomness the model uses when choosing each next token. Near zero, it almost always picks the most likely token, giving repeatable, focused output — the right choice for classification, extraction, or anything you want to be consistent.

Higher temperatures let it sample less likely tokens, producing more varied and creative results at the cost of reliability. It is a decoding setting, not a prompt instruction: writing 'be creative' in the prompt is not the same as raising the temperature, and for structured tasks you usually want it low regardless of wording.