Qwen 3.6's Reasoning Mode: A Silent Killer for Structured Output Agents

Qwen 3.6 ships with reasoning mode enabled by default, injecting <think> tags with extended chain-of-thought reasoning into every response. For developers building agents that require clean JSON output, this creates silent parsing failures and wastes token budgets unpredictably. A simple parameter flag in apply_chat_template solves the issue instantly.