After extensively testing most of the domestic large models on the market, especially in Agent scenarios, here are my personal favorites: Xiaomi @XiaomiMiMo and @deepseek_ai. My two Hermes setups are powered by mimo-v2.5 and deepseek-v4-pro. I only switch to GPT-5.5 for heavy coding or deep research scenarios. These two models are truly generous in terms of token capacity. Currently, the cache hit rate for the deepseek model I’m using is insanely high—inputting a million tokens (cache hit) costs just ¥0.1. With the current discount, it’s only ¥0.025. This means a complete Agent call often costs just a few cents, or even fractions of a cent. At this price point, for daily tasks that are tedious, token-intensive, and don’t demand extreme quality, the cost-performance ratio is undeniably high. One unexpected benefit: the emotional nuance in language expression. People who are sensitive to text will get this—Chinese large models naturally convey a smooth sense of tone and subtlety in conversations. On the other hand, using English-trained models to speak Chinese always feels a bit stiff, like there’s a layer of gauze in between.