Qwen 3.8 27B xhigh thinking burns 23.8K tokens on simple prompt

A Reddit user running a local Qwen 3.8 27B (UD_Q4_XL) compared its thinking-effort settings against DeepSeek V4 Flash, ChatGPT Free, and Claude Opus 5. The prompt was simple: "Write a simple html CARD about the benefits of eating apple. paste the code here." All models except Qwen's 'xhigh' effort treated it as trivial, responding with minimal reasoning and around 800-1,000 output tokens. Qwen 3.8 27B in 'xhigh' mode, however, generated 23.8K tokens of deep reasoning before producing a far more polished, feature-rich HTML card—much better than anything else in the test.

The user then reran Qwen's 'medium' mode with a more detailed prompt specifying an 'orchard notes' design, pastel colors, stats, and an interactive layout. It followed instructions precisely with only 3.3K output tokens, achieving quality close to the xhigh result. The takeaway: xhigh can overthink even simple requests, costing massive tokens, but the output quality scales accordingly. Medium mode trades that depth for speed and lower token usage, making it a solid default for low-throughput setups. For users with limited VRAM or API budgets, prompt specificity can compensate for lower thinking effort—a practical insight for anyone deploying local LLMs.