Logits are raw output scores from a model representing the unnormalized likelihood of each token. Token probabilities convert these logits into normalized values, indicating the chance each token will be selected next.
The model generates logits for every possible token in its vocabulary. These logits are passed through a softmax function, normalizing them into probabilities that sum to one. The token with the highest probability is typically chosen for output.
Accurate token probabilities enable better control over AI output quality and diversity. For product managers, this affects user experience by improving response relevance, impacts cost and latency by influencing sampling strategies, and supports scalable deployment by managing computational resources efficiently.