It is basically a sort of dynamic top_k which passes all tokens that score at least min_p*best. It works well for creative output because it is more flexible when there are many possible continuations of similar probability.

More info: https://github.com/ggerganov/llama.cpp/pull/3841