Logits
The raw scores a model gives every token in its vocabulary before they are turned into probabilities.
At the end of the last layer, a model multiplies the final hidden state by a large matrix to get one score, a logit, for every token in its vocabulary, often more than 100,000 numbers. Higher means more likely next. Softmax converts the logits into probabilities, and a sampling method picks a token. Some APIs return log-probabilities, which developers use to measure how confident a model is, to detect when it is guessing, or to force it to choose between a few allowed answers.