Open Models Almanac
Methodsinference & servingdecoding strategy

implementation detail · filed under inference & serving

Presence penalty tuning

Adjusts the presence penalty to reduce repeated or endless output.

The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See decoding strategy for the methods filed beside it.

Also called presence_penalty.

sources
2
model
1
lab adopt it
1
strongest
used

Documented in

Alibaba (Qwen)

Evidence

3 quotes from 2 sources, one card per model, each opening with its strongest.

How sources treat it
optional 2used 1

Qwen3.5-397B-A17B

Alibaba (Qwen)adopted2 sources

For supported frameworks, you can adjust the presence_penaltyparameter between 0 and 2 to reduce endless repetitions.

usedinference servingin Qwen3.5-122B-A10BQwen Team
2 more from this model

you can adjust the presence_penaltyparameter between 0 and 2 to reduce endless repetions.

optionalinference servingin Qwen3.5-35B-A3BQwen Team

For supported frameworks, you can adjust the presence_penalty parameter between 0 and 2 to reduce endless repetitions.

optionalinference servingin Qwen3.5-35B-A3BQwen Team
↑