Qwen3.5 dialogue turns courtly, users gripe

This looks less like a broken model and more like a model whose default post-training leans hard toward reasoning, tool use, and formal answer style. If you want casual, contemporary dialogue, you probably need to fight the defaults instead of assuming prompt text alone will fix it.

• –The official Qwen3.5 docs say thinking mode is on by default, and they recommend `enable_thinking: False` or the equivalent chat-template setting for direct responses.
• –Qwen3.5 is positioned as a native vision-language model with agentic capabilities, so its “safe” conversational style is optimized for precision and structure, not punchy dialogue.
• –Community chatter around Qwen3.5 lines up with this: people keep describing it as overthinking or sounding too polished in casual chat.
• –For your use case, the likely fixes are format-focused: disable thinking, lower formality with few-shot examples, and tune sampling for shorter, more human lines.
• –If the output is for video prompt generation plus dialogue, Qwen3.5 may still be a good backbone, but it seems better at scaffolding than at voicey creative writing out of the box.