Logits are the raw, unnormalized scores a neural network produces for each possible output before they are turned into probabilities. In a language model, the final layer emits one logit for every token in the vocabulary, and a softmax function converts those numbers into the probabilities the model uses to choose the next token. Higher logits mean the model favors that option, but on their own they are just scores, not probabilities.
Also known as raw model scores, pre-softmax outputs
When a model makes a prediction, its last layer outputs a vector of logits, one number per class or token. These values can be any real number, positive or negative, and they are not directly comparable as probabilities. To turn them into a probability distribution that sums to one, the model applies a softmax function, which exponentiates and normalizes the logits so the largest logits become the highest probabilities.
Logits matter because most of the controls over how a model generates text act on them. Temperature scales the logits before softmax to make the output more or less random. Techniques that ban or boost certain tokens, or that bias the model toward a format, do so by adjusting logits. Reading the logits also reveals how confident a model is and how close the runner-up options were.
For agencies, logits are not something you handle day to day, but they explain the dials that change how AI tools behave.
They are what the model’s settings actually move. When a tool offers a creativity or randomness control, it is usually scaling logits under the hood, so understanding them demystifies why the same prompt can give safe or wild results.
They underpin reliability tricks. Forcing a model to return valid options, a fixed format, or a constrained choice often works by editing logits, which is how vendors make AI output predictable enough to drop into a workflow.
They are a signal of confidence. Teams building or evaluating AI tools can read logits to see how sure a model was, which helps decide when output needs a human check before it reaches a client.
An agency builds a tool that tags incoming creative assets by type. Early on, it returns messy, inconsistent labels. The engineer constrains generation at the logits level so the model can only choose from the approved list of tags, and lowers the temperature so it stops inventing new ones. The same model now returns clean, predictable labels the team can trust. The fix was not a better prompt; it was shaping the logits the model was allowed to act on.
The automations and agents module of the workshop teaches you how to build AI workflows that compress the busywork without taking the craft out of the studio.