Why Is Qwen 3.8 27B So Slow? One Default Setting.

#Qwen #LocalLLM #AIEngineering Simon Willison ran one prompt on a seventeen gigabyte model that fits on a laptop. Draw an SVG of a circle. It took **twenty-one minutes** and **22,276 tokens of internal reasoning** to produce 3,223 tokens of finished work — and what came back was a "geometric circle study" with concentric guides, tick marks and ambient motion. Not a circle. The model is genuinely excellent. The problem is one line of configuration. Qwen ships **Qwen 3.8 27B** with `reasoning_effort` set to **xhigh**, the level its own documentation reserves for "complex tasks demanding thorough analysis" — applied to every prompt, in every quantised build. Willison's verdict on that choice is one word: hilarious. This breakdown works through what the default costs and when it is worth paying: - the arithmetic: 22,276 reasoning tokens against LM Studio's 8,192-token default context window, which the model exhausted before answering anything - the same prompt with reasoning off — 3,715 tokens in 137 seconds, twenty-one minutes down to just over two - and what that speed costs: with reasoning on, the drawing had a correct frame, legs on both sides, wings on the handlebars. With it off, the frame breaks and the feet miss the pedals - where it is genuinely strong: JSON bounding boxes returned clean and landed on the birds, no parsing gymnastics - the one-shot build that **nearly** worked with reasoning off — the interface came out solid and the boxes rendered in the wrong place, which is the failure mode you ship by accident - driving a real coding agent against the Datasette codebase, and writing plus testing a Python script unprompted - the throughput ceiling: 15–30 tokens per second locally against 74 and 184 for hosted models, and why a dense model is memory-bandwidth bound - multi-token prediction, already inside the file, measured around **72% faster** with one serving flag The rule worth keeping: **reasoning effort is a budget you set per job, not a setting you leave alone.** Run it low or off for extraction, classification and bounding boxes, where the answer has one correct shape and the thinking buys you nothing. Spend it on one-shot builds and anything where being wrong is expensive to notice. One caution, and it comes from Willison rather than from this channel: Qwen's published benchmarks are self-reported. No independent numbers exist yet. ### Chapters ```text 0:00 Twenty-one minutes for one drawing 0:36 What the default costs, and the catch 1:06 Three documented levels, one shipped 1:43 It ate the context window 2:13 Asked for a circle, got a study 2:48 Reasoning off: twenty-one minutes becomes two 3:21 What switching it off costs 3:51 Where it is very good: vision 4:23 One prompt, a working tool 4:52 And the same tool broke without it 5:18 Driving a real coding agent 5:49 The real constraint is speed 6:22 Why dense models are slow here 6:50 Seventy-two percent, from a serving flag 7:21 The benchmarks are self-reported 7:50 Set it per task 8:19 Who should run it 8:46 What actually changed this year ``` ### Primary sources - Simon Willison — Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things: https://simonwillison.net/2026/Aug/16/qwen-38-27b/ - Qwen3.8-27B model card (vendor): https://huggingface.co/Qwen/Qwen3.8-27B Every figure in this video comes from those two pages. Nothing was benchmarked by this channel, and the description says so.