Limits token sampling to the top K most likely tokens. Higher values increase output diversity. Default: 40