Number of top most-likely candidates, between 0 and 500, from which the model chooses the next token in the sequence.