Qwen3.5 reasoning budget stumps vLLM, SGLang

Hot take: this looks less like a model bug and more like a docs-and-runtime mismatch. Qwen3.5 has explicit thinking controls, but vLLM and SGLang expose them as parser/template flags rather than a portable “reasoning-budget” primitive, so it’s easy to misconfigure.

• –Qwen’s repo says `enable_thinking=False` is the strict switch, and `/think` / `/no_think` can toggle behavior turn by turn.
• –vLLM documents the `qwen3` reasoning parser and server defaults via `--default-chat-template-kwargs '{"enable_thinking": false}'`, with request-level kwargs overriding the default.
• –SGLang’s Qwen 3.5 guide shows `--reasoning-parser qwen3` and `--tool-call-parser qwen3_coder`, but no separate budget knob in the basic launch flow.
• –If the engine still allows free-form thinking, prompt tweaks alone probably won’t stop the model from burning tokens.
• –For production use, this really wants a single compatibility matrix for “thinking on/off,” “budget cap,” and “framework support.”