Qwen 3.8 27B defaults to xhigh reasoning effort consuming 22,276 tokens on simple SVG tasks

Alibaba released Qwen 3.8 27B on 16 August 2026 under Apache 2.0 with vision support. The model preserves the xhigh default in the official 17GB Q4_K_M GGUF build. Simon Willison tested it on M5 Max and DGX Spark hardware via LM Studio with context extended to 262,144 tokens after the 8,192 default proved insufficient.

Self-reported benchmarks show gains over Qwen 3.6 27B and Qwen 3.7-Plus. Actual runs reveal the cost: the circle SVG prompt triggered a multi-paragraph internal trace on concentric rings, gradients and reduced-motion handling before emitting output. Token counts scale directly with the fixed xhigh setting regardless of task complexity.

Overthinking at xhigh reduces effective throughput on consumer hardware and raises inference cost without accuracy gains on trivial prompts. This pattern echoes documented chain-of-thought overhead in other models where fixed high reasoning budgets degrade utility on latency-sensitive or low-complexity workloads. The default conflicts with deployment constraints documented in the GGUF quantization notes.

LM Studio and downstream hosts will likely expose reasoning_effort as a first-class toggle within six months. Independent benchmarks from OpenRouter runs of the 2.4T variant already demonstrate faster paths when the parameter is lowered.