Using the thinking 0121 model for translation often results in a mix of original text and translated text

As shown in the screenshot, I used Chinese prompts to ask the API to translate a Chinese novel paragraph by paragraph into English, but the first few paragraphs were not translated (sometimes there were even cases where a sentence was translated halfway, with the other half remaining in Chinese). Then it started to work normally. I’m not sure what caused this.

While this reply may prove unhelpful if you specifically require gemini-2.0-flash-thinking-exp-01-21 or if this is merely a bug report, I see no reason to employ thinking models as if analyzing a conspiracy or equation. gemini-exp-1206 excels in translation, based on my experience (though errors and occasional foreign text insertions may occur, they are rare). I used it to translate a novel from English to Arabic, requesting the use of classical language; its eloquence surpassed that of some native speakers.

Yes, I’ve noticed that Gemini-1206 is the strongest model in its series for translation. I’m working on an application specifically designed for translating long texts, like novels, so I’m testing the capabilities of various models. I’d also like to ask: if I create such a program—capable of translating 200,000 characters in under two minutes, supporting any language a large language model can handle—would you be willing to pay for its use? Pricing would be per character, and there would be specialized prompt optimization to ensure literary quality and coherence in the translations. I’m considering pricing it at 4-5 times the cost of Gemini. Do you think there would be a significant global market for such a service?