MiniMax has identified a 4.9% degradation in the token vocabulary of its M2 series large model, as detailed in a recent technical blog. The issue was discovered during an investigation into the model's inability to output the name "Ma Jiaqi." The root cause was traced to the tokenizer merging "Jiaqi" into a single token during training, which led to its degradation due to low-frequency usage in post-training data. A comprehensive scan revealed that Japanese tokens were most affected, with 29.7% showing significant degradation, compared to lower rates in Korean, Russian, Chinese, and English. The degradation caused language mixing and token forgetting, notably inserting Russian or Korean characters into Japanese dialogues. MiniMax addressed the issue by training the model on a synthetic dataset, significantly reducing errors and improving token stability.