Temperature & Sampling Parameters

Temperature controls randomness in model output. Lower temperature (0.0-0.3) produces deterministic, conservative responses — the model picks the most likely token at each step. Higher temperature (0.7-1.0) produces more varied, creative responses — the model samples from a broader probability distribution. For code generation, lower temperature is almost always correct. Code has right answers and wrong answers; creativity in token selection means bugs. For brainstorming, higher temperature produces more diverse ideas.

Top-p (nucleus sampling) is temperature's companion parameter. While temperature scales the probability distribution, top-p truncates it — only tokens in the top p% of probability mass are considered. A top-p of 0.95 means the model considers tokens that collectively make up 95% of the probability, ignoring the long tail of unlikely tokens. For code, top-p at 0.9-0.95 paired with temperature at 0.1-0.3 produces reliable, correct output. For creative tasks, top-p at 0.95-1.0 with temperature at 0.7-0.9 produces diverse output.

{
  "profiles": {
    "precise": {
      "model": "gemini-2.5-pro",
      "temperature": 0.1,
      "topP": 0.9,
      "thinkingBudget": "auto"
    },
    "creative": {
      "model": "gemini-2.5-pro",
      "temperature": 0.8,
      "topP": 0.95,
      "thinkingBudget": "auto"
    },
    "deterministic": {
      "model": "gemini-2.5-pro",
      "temperature": 0.0,
      "topP": 0.9,
      "thinkingBudget": "auto"
    }
  }
}