Saving on LLM: Chinese Model Unexpectedly Started Writing in Chinese Characters
A developer of a bilingual news bot decided to switch to a cheaper AI model: Ling 2.6 Flash by Inclusion AI costs 8 times less than DeepSeek V4 Flash and runs 4 times faster. However, after an hour of use, Chinese characters appeared in the feed: the model periodically reverted to its native Chinese language, especially on longer generations. The problem was solved by adding a Chinese character detector with a re-query to the clean model.
Tencent
DeepSeek
Habr — хаб ML27.07 · 04:03