Temperature
A setting that controls how adventurous sampling is. Low temperature gives safe, predictable text; high gives varied text.
Temperature divides the logits before softmax. At a temperature below one, the gap between likely and unlikely tokens widens and the model almost always picks its top choice; at zero it becomes greedy decoding. Above one, the probabilities flatten and rarer tokens are chosen more often, which can help brainstorming but raises the chance of errors. Many APIs default to about 0.7 to 1. Some reasoning models fix temperature themselves and ignore the user's setting.