Temperature

What it is

IBM describes temperature as a parameter that controls the randomness of text an LLM generates. At low temperature the model mostly picks the most likely next token. At higher temperature, less likely tokens get chosen more often.

How it works

1. Set temperature in the API request. Apps usually hide it.
2. Low values sharpen the probability distribution and high values flatten it. Google’s prompting guide says lower temperatures suit prompts that need a more deterministic response, and higher ones can give more diverse or creative results.
3. Combine it with clear instructions. Temperature changes how much the output varies, not what the model knows.

Newer models (checked 2026-10-17). Google strongly recommends keeping temperature at its default of 1.0 for all Gemini 3 models, because their reasoning is optimized for that setting. Check each vendor’s model guide before tuning.

Benefits

• More consistent outputs for data tasks, on models where tuning is recommended.
• More diverse ideas for creative work.
• An easy lever when outputs feel repetitive or erratic.

Limits and risks

• Low temperature doesn’t prevent hallucination.
• Allowed ranges, defaults and recommendations differ by vendor and model.
• Even at zero, outputs may not be perfectly identical.
• For consistent formats, structured outputs are more reliable than temperature.