I had Gemini 2.5 PRO, Sonnet 4.5, and MiniMax-M2 take turns optimizing the translation prompts for Qwen-4B. After a whole day of intense work, they now consider me a genius 🤣. I switched LM Studio to Qwen3-4b-2507 8bit. 4.29 GB is enough for M4, the speed is acceptable, and it retains the performance of the original model to the greatest extent possible. It feels perfect.
