Model mixes Japanese characters (検索) in Portuguese responses

[BunpGhost]

Model mixes Japanese characters (検索) in Portuguese responses

Model: MiniMax-M2.7 (minimax-portal/MiniMax-M2.7)
Language: Portuguese (PT-PT)
Channel: OpenClaw AI assistant via Discord

Problem

When responding in Portuguese, the model frequently inserts Japanese characters into the response without any apparent reason. This is not an isolated incident — it happens multiple times per conversation.

Expected behavior

The model should respond entirely in Portuguese (PT-PT) without mixing characters from other languages.

Actual behavior

Example from conversation:

"Agora com検索 inteligente — antes só encontravas por keyword exata, agora encontra por significado."

The word 検索 means "search" in Japanese and appears mixed into a Portuguese sentence without any context or justification.

Frequency

Happens multiple times per session. Not triggered by any specific prompt pattern — appears randomly during normal conversation.

Steps to reproduce

1. Use MiniMax-M2.7 model
2. Have a conversation in Portuguese (PT-PT)
3. Wait for responses about search functionality or similar topics
4. Japanese characters appear unexpectedly

Additional context

• This behavior is inconsistent — most responses are fine, but roughly 1 in 5-10 responses contains mixed Japanese
• The model clearly knows Portuguese, as most responses are fully in PT-PT
• No encoding issues detected on the client side (Discord/web interface)
• Other models tested: Codex (GPT-5.4) — no issues observed

Possible cause

Potential issue with the model's tokeniser or language detection mechanism when handling multilingual content, or possible contamination from training data that includes Japanese text.

[ximiximi423]

Thanks for your feedback. Could you please share the specific request and response content to help us look into it?

[BunpGhost]

Additional example with specific request/response pair:

Request:
"Explica como funciona a pesquisa de memória no OpenClaw"

Response (containing the bug):
"A pesquisa de memória usa检索 inteligente — combina busca vectorial com matching semântico para encontrar resultados mesmo quando não usas exactamente as mesmas palavras."

The character 检索 (Chinese for "search/retrieve") appears mid-sentence with no contextual justification. This is the same phenomenon as the Japanese 検索 example in the original report — the model is inserting CJK characters into otherwise clean Portuguese text.

Scope clarification: the issue is not limited to Japanese. The bug affects Chinese characters (检索) and Japanese characters (検索) equally, appearing unpredictably in roughly 1 in 5–10 responses during normal PT-PT conversation. The root cause appears to be in the model's tokenizer or multilingual decoding logic when operating in a Portuguese context.

---

Note: the issue title says "Japanese characters" but the actual phenomenon is broader — CJK characters (Chinese/Japanese) mixed into Portuguese. Consider updating the title to reflect this.