During testing of Xiaomi’s large model MiMo-v2-flash, users observed significant variations in model output depending on network environments. When users connected directly to L-site using custom DNS, the model accurately identified itself as “MiMo”; however, other users enabling search mode found the model自称 “Hajimi”. Through systematic questioning, users inferred that the LLM service provider might determine user geographical locations based on IP addresses, thereby delivering differentiated services and even implementing precise “capability degradation”. This phenomenon not only involves technical aspects but also sparks widespread discussion about AI service fairness and transparency. The article presents actual case studies demonstrating behavioral differences in AI models during global deployment, offering valuable insights for readers focused on cutting-edge AI technology.
