loading
loading
The raw, unnormalized scores a model outputs for every possible next token before they become probabilities.
At each step the model produces one logit per token in its vocabulary — a big vector of raw scores saying how much it 'likes' each candidate. Running them through softmax turns those scores into a probability distribution you can sample from. Logits are the layer where temperature and top-p/top-k actually operate: you reshape the logits, then pick. If you ever do constrained decoding or grammar enforcement, you're masking logits.
Plainly
Think of Logits as the brain part that guesses or decides. The raw, unnormalized scores a model outputs for every possible next token before they become probabilities.
In practice
Use it when model choice, prompts, latency, cost, or quality affect the product result. In practice, define the owner, input, output, and failure mode before you rely on it.