[Model BadCase]: Korean responses mixed with Chinese characters (한자) - cannot output pure Korean #100
Description
Basic Information - Models Used
MiniMax M2.7
Information about environment and deployment
Basic Information
- Models Used: MiniMax-M2.7 (and MiniMax-M2.5)
- Information about environment and deployment: Official API
- Channel: OpenClaw (Telegram bot)
Description
When using MiniMax-M2 for Korean language tasks, the model frequently mixes Chinese characters (한자/漢字) into Korean responses despite explicit instructions to output Korean only.
Steps to reproduce
- Set system/user prompt: "한국어만 사용" / "Korean only"
- Request any response in Korean (e.g., weather summary, simple explanation)
- Observe Chinese characters (한자) appearing in the output
Expected behavior
- All output should be in Korean only
- No Chinese characters (漢字) or Japanese characters should appear
- Example: "컴퓨터" not "電腦", "자동차" not "汽車"
Actual behavior
- Model outputs Korean text mixed with Chinese characters
- Example recent output: "설정 파일을 확인했어요. 현재 SOUL.md를 시스템 프롬프트로 사용하고 있는데, 여기에 규칙을 추가하겠습니다."
- Words like "설정", "파일", "확인" sometimes rendered with embedded Chinese characters
- The issue persists despite explicit system prompt instructions
Severity
- High for Korean-speaking users - makes responses difficult to read
- Affects all Korean language use cases
Related reports
- Similar report on X (Twitter): A Korean user reported the same issue with MiniMax-M2.5
- GitHub Issue [Model BadCase]: The model always confuses between Chinese and English Punctuations #55: Chinese/English punctuation confusion suggests broader character handling issues
Additional context
This appears to be a fundamental limitation in how the model handles multilingual (Korean + Chinese) text generation. Users requesting Korean-only output cannot achieve it reliably.
Description
Steps to reproduce
The bug can be reproduced with the following steps: