When generating the list of possible tokens, the model considers the smallest number of tokens with a probability that add up to be greater top_p.

#ollama/parameters