min-p sampling

Min-p sampling is a newer, confidence-relative filter. Rather than a fixed count (top-k) or a cumulative mass (top-p), it sets the bar relative to the single most likely token. You choose a fraction — say 0.05 — and keep only tokens whose probability is at least that fraction of the top token's probability. If the leading token sits at 0.6, the threshold is 0.03, and anything below that is dropped before sampling.

This adapts gracefully to how peaked the distribution is. When one token towers over the rest, the threshold rises with it and the pool stays tight, protecting quality even at high temperature; when the field is flat and many tokens are comparably likely, the bar is low and lots of options survive, keeping diversity. Fans favour min-p because it lets you crank temperature for creativity without the gibberish that pure top-p can let slip through. It is increasingly offered alongside top-k and top-p.