Top-p

After scoring its vocabulary, a model holds a probability for every possible next token. Top-p sorts them from most to least likely, adds them up from the top until the total reaches p, and discards everything below that line. The tokens that remain are rescaled so their probabilities sum to one, and the choice is made among them. The size of the group is not fixed: when the model is confident, a single token may already pass the threshold, and when many continuations are plausible the group grows to include dozens. That adaptive behavior is why top-p became popular next to top-k, a simpler cut that always keeps the same number of candidates whatever the situation.

Think of a restaurant menu that lists only the dishes most guests order, until they cover nine orders in ten. On a quiet evening with a clear favorite the menu is short; on a busy night with varied tastes it is long. A top-p of 1 keeps every token and so changes nothing, while a value near 0.1 keeps almost only the best choice and makes the output close to deterministic. For conversation and writing, values between 0.8 and 1 are typical, and lowering the figure trims the rare, odd words that can make a text drift off course. Many chat interfaces hide this control, yet it remains a handy lever when an assistant sounds repetitive or wanders.

Top-p acts at the same stage as temperature, and the two interact. Temperature reshapes the probabilities first, then top-p cuts the tail of the reshaped distribution. Most guides therefore suggest changing one and leaving the other at its default, so you can tell which of them caused a difference. Neither setting makes an answer truer: they only shape how adventurous the wording is. Some reasoning models fix these parameters or ignore them, and providers differ in which sampling controls they expose, so a model’s own documentation has the final word.

On NeverNo, the API accepts top_p in a chat request, next to temperature, stop and seed. They are passed to the models that take them and left out for the others, and the models list shows which parameters each model supports, so check it before relying on one. Qwen 3.5 35B, a free private model from Alibaba that reads images and lets you choose whether it reasons, is one of the open models in the catalog.

The page about uncensored AI chat shows how NeverNo answers frankly while keeping your history in your browser. Sampling settings change the style of an answer, never the hard limits: nothing involving minors, nothing about real people without their consent, and no real help to seriously harm others.