Korean tool calls stall on Qwen3.5

Hot take: this is probably a tooling/prompt-format problem, not “Korean is impossible.”

• –Qwen3.5 is documented to default to thinking mode, and its tool-use path depends on the serving stack handling the model’s output correctly.
• –The fact that English works but Korean hangs suggests the prompt/template or parser is less robust outside English, not that the model cannot reason in Korean.
• –A stronger system prompt that forces a fixed tool-call format can help, but if the backend/parser is brittle you may still need serving-side changes.
• –If the model is being run with the wrong reasoning or tool-call parser, it can look like “I will read the file now:” and then stall instead of emitting the actual function call.
• –Inference: the most likely fix is a combination of tighter tool-call instructions, explicit language constraints for the tool-call segment, and verifying the inference backend supports Qwen3.5 tool calls correctly.