Token Probabilities Explained: Logits, Softmax, and Temperature
By Rui Barreira · Last updated: 13 June 2026
LLMs generate text one token at a time. For each step, the model produces scores for candidate tokens. brevio Token Probability Visualizer shows how those scores become probabilities through softmax.
Logits
A logit is an unnormalized token score. Higher logits are more likely to be selected, but logits are not probabilities until they are normalized.
Softmax
Softmax exponentiates each logit and divides by the total across all candidates. The result is a probability distribution where all values add up to 1.
Temperature
Temperature scales logits before softmax. Lower temperature makes the most likely token more dominant. Higher temperature spreads probability across more candidates, increasing variety but also increasing risk.
Use Cases
This is useful for understanding why deterministic settings produce repetitive output, why high-temperature settings can drift, and why small logit differences can become large probability differences.
Frequently Asked Questions
- What is a logit?
- A logit is an unnormalized score for a candidate token. Higher logits become higher probabilities after softmax.
- What does softmax do?
- Softmax exponentiates the logits and normalizes them so all candidate probabilities add up to 1.
- How does temperature affect output?
- Lower temperature makes the highest-logit token more dominant. Higher temperature spreads probability across more candidates.
- Does this tool use a real model?
- No. It is an educational calculator for candidate logits you enter manually.