Qwen3.5 users trim thinking, token bloat

Qwen3.5’s “overthinking” is the real story here: if you want shorter answers, runtime controls usually matter more than clever wording. The thread is useful less as a prompt showcase and more as a reminder that token efficiency is a stack problem, not a single prompt trick.

• –Qwen’s own docs and GitHub discussions point to `enable_thinking=False` or `/no_think` as the most direct way to cut reasoning output when your deployment stack supports it.
• –A minimal, directive system prompt tends to outperform long style instructions, because prompt bloat eats context before it ever saves tokens.
• –Generation settings like `max_tokens`, temperature, top-p, and presence penalty can materially change how chatty the model gets.
• –For repeatable brevity, external constraints such as response schemas, stop sequences, and output templates are more reliable than asking the model to self-police.
• –The thread captures a common local-LLM tradeoff: Qwen3.5 can be capable, but efficiency depends heavily on prompt template discipline and serving configuration.