Temperature

What is temperature?

Every time a model generates the next word, it's really drawing lots from a pool of candidates, each with its own probability. Temperature is the knob that sets how hot or cold that draw is: does the model reliably pick the most likely word, or does it sometimes gamble on a long shot?

How does it work?

Lower = more conservative
At low temperature the model almost always takes the safest word. Output is steady and predictable — good for code, translation and factual questions.
Higher = bolder
High temperature gives the long shots a real chance. Output gets more surprising and creative — good for stories and slogans, but also more prone to nonsense.

Which values to pick?

Near 0
Almost always the best option, results are reproducible — great for precise tasks.
0.7 to 1.0
The sweet spot for everyday chat and general writing.
Above 1.0
Free-for-all mode, good for brainstorming — just plan to filter by hand.

What it isn't

Temperature is not a “smartness” switch. Turning it up doesn't make the model smarter, just more talkative; turning it down doesn't make it dumber, just more restrained. It governs randomness, not ability.

Bottom line: temperature is the AI's “courage dial” — turn it down for reliability, up for surprise.