Prompt
"The cat sat on the ___"
Most probable Less probable
Bar chart of token probabilities. High temperature flattens the distribution; low temperature sharpens it.
Temperature 1.0
← focused random →

The model holds a distribution like this at every single generation step — over tens of thousands of tokens. Temperature controls the tradeoff between predictability and variety. At 0.1, the top token wins almost every time. At 2.5, anything can happen.