I benchmarked every reasoning level on Qwen 3.8-27B

Yesterday the 4-bit build won my quant showdown with thinking turned off. But thinking is Qwen 3.8's default mode, so overnight I ran the same exam at every position of the reasoning dial: off, low, medium, and xhigh (the factory default). Same 17.9 GB file, same four one-shot tests, one variable. Then I went looking for how the dial actually works, and that turned out to be the best part.

The scoreboard nobody predicted

Every setting won exactly one test. Medium took the Tetris clone (2 minutes 58 seconds against the default's 52 minutes). Xhigh took the volcano sim, it simply looked better and worked. Low took the spreadsheet trap. And reasoning-off took the drawing test, as the only run that both rendered AND signed its own name. Four settings, four wins: the dial is a per-task decision, which is presumably why it is a dial.

The default is expensive

At xhigh, the whole exam took 2 hours 36 minutes, and over 2 hours of that was pure thinking: on the Tetris test alone it wrote 122,000 characters of deliberation and 53,491 output tokens. Medium finished the same exam in 22 minutes. Thinking tokens generate at the same ~18 tok/s as answer tokens, so every minute of thinking is a minute you wait.

Medium beating low was not a fluke, Qwen called it

Their model card warns that lower reasoning effort "does not always reduce overall task completion time... it can also lead to insufficient analysis, more failures, and repeated retries." That is exactly what my numbers show: low thought more than medium on three of four tests and finished slower overall. My new default is medium, and I am daily-driving it in agent harnesses for the next week.

The level that does not exist

Unsloth's docs and app list a fourth level, "none". I ran it as a fifth contestant, and every request failed with the model's own template throwing the error: "Unexpected reasoning effort none. Supported types are xhigh (default), medium, and low." The off switch is real, it is just a different parameter (enable_thinking: false), and the docs listing "none" as an effort level is a bug I got to discover on camera.

The signature saga finale

My drawing test asks the model to sign the artwork with its own name. The same file, at four settings, signed:

Three borrowed identities in one night, all verified in the raw outputs on my machine. Open models are trained on frontier-model text, and it rarely shows up this visibly.

The reveal: the dial is three sentences

I extracted the chat template from inside the GGUF file. The reasoning levels are not compute budgets or special decoding modes. Each level just injects a different preamble into your prompt, server-side. Xhigh injects a paragraph asking the model to "think carefully through the task, validate key assumptions, consider plausible alternatives." Low injects "keep your thinking brief and focused." And medium injects NOTHING, there is no medium branch in the template, it is thinking with zero coaching. A few injected words are the entire difference between the 22-minute exam and the 156-minute one.

Every number, prompt, output, thinking transcript, and the extracted template itself: the full results page.

Day 30 of 30

This is the last daily video. Thirty days ago I committed to creating in public every single day, and today the streak completes: a benchmark franchise, a newsletter, these letters, and more learning per week than any month I can remember. The channel continues, just not daily. To everyone who commented, corrected my benchmarks, and subscribed along the way: thank you. The next video comes when there is something worth showing you, and the way this field moves, that will not take long.

My rig

Sources

• The exact file tested: Qwen 3.8-27B UD-Q4_K_XL GGUF
• The modes, samplers, and the multi-turn caveat: official Qwen 3.8 model card
• The docs listing "none": Unsloth's Qwen 3.8 page
• The series: the BF16 showdown · the quant ladder

I am a self-taught engineer figuring out tech: AI, Data Privacy, Linux, and everything in between. New videos ongoing.