Han Xiao(@hxiao):if you find salad words or CJK bleeding in reasoning traces of qwen3.8-flash-next on high thinking, remember to set presence_penalty=0; for non-thinking-mode which many users do they do presence_penalty=1.5. bc presence penalty assumes repetition is a defect, which works for the final answer. but for a reasoning trace, repetition is the model keeping hold of the problem, so the high penalty not helping (as we do need to restate the goal, wait-gaslighting in reasoning trace).

2026.09.09 23:43

if you find salad words or CJK bleeding in reasoning traces of qwen3.8-flash-next on high thinking, remember to set presence_penalty=0; for non-thinking-mode which many users do they do presence_penalty=1.5. bc presence penalty assumes repetition is a defect, which works for the final answer. but for a reasoning trace, repetition is the model keeping hold of the problem, so the high penalty not helping (as we do need to restate the goal, wait-gaslighting in reasoning trace).