The AI Handbook Open in the app →

Temperature

Inference & Prompting · Beginner · 4 min read

What is it?

Temperature is a setting that controls how random a language model's output is — low temperature makes it focused and predictable, high temperature makes it varied and creative.

Explain like I'm 5

It's a creativity dial. Turn it down and the model plays it safe with the most likely words; turn it up and it takes more chances.

Why was it created?

A model produces a probability for each possible next token. Temperature was introduced as a simple knob to trade off between reliable, repeatable answers and diverse, creative ones.

Where is it used?

  • Chat and completion APIs
  • Creative writing tools
  • Deterministic extraction tasks
  • Sampling configuration

Why should developers care?

It's one of the most common settings you'll tune when building with LLMs, and it directly affects reliability, so understanding it saves a lot of trial and error.

How does it work?

Before picking the next token, the model has a probability for each option. Temperature rescales those probabilities: low values sharpen them (the top choice dominates), high values flatten them (less-likely tokens get a real chance). Temperature 0 makes it (nearly) deterministic.

Real-world example

For extracting a date from text you set temperature near 0 for consistency; for brainstorming taglines you raise it to get varied ideas.

Common use cases

  • Deterministic outputs (temp ~0)
  • Creative or varied output (higher temp)
  • Reducing repetition
  • A/B testing prompt behavior

Advantages

  • Simple, single knob
  • Trades reliability for creativity on demand
  • Supported by virtually every model API
  • Cheap to experiment with

Disadvantages

  • High values increase errors and drift
  • Interacts with other sampling settings (top-p)
  • Not a quality control by itself
  • Same value behaves differently across models

When should you use it?

Lower it for factual, structured, or repeatable tasks; raise it for brainstorming and creative variety.

When should you avoid it?

Don't rely on temperature alone to fix quality or accuracy — grounding and prompting matter more.

Alternatives

Top-p (nucleus) samplingTop-k samplingGreedy decoding

Related terms

TokenLarge Language ModelInferencePrompt Engineering

Interview questions

Beginner

  • What does raising temperature do to model output?
  • What temperature would you use for consistent extraction?

Intermediate

  • How does temperature reshape the token probability distribution?
  • How do temperature and top-p interact?

Senior

  • Why can temperature 0 still produce different outputs across runs or hardware?
  • When would you prefer top-p over temperature?

Common misconceptions

  • "Higher temperature is smarter" — it's just more random, often less accurate.
  • "Temperature 0 is fully deterministic everywhere" — implementation and hardware details can still cause variation.

Fun facts

  • The name comes from the physics of the softmax function used to sample tokens.
  • Many production extraction pipelines run at temperature 0.

Timeline

  • 2010s — Softmax temperature used across neural network sampling
  • 2020s — Becomes a standard knob in LLM APIs

Learning resources

Quick summary

Temperature is a knob for output randomness: low values give focused, repeatable answers; high values give varied, creative ones.

Cheat sheet

  • Creativity/randomness dial
  • Low = focused, high = varied
  • ~0 for extraction, higher for ideas
  • Pairs with top-p sampling

If you remember only one thing

Temperature is the randomness dial — turn it down for reliable answers, up for creative variety.