Qwen 3.8 27B: Q4_K_M matches BF16, xhigh is what to change

Two write-ups on Qwen 3.8 27B came out a few weeks apart. Read together, they say the quantisation is the easy choice and the reasoning effort is the one that bites.

The benchmark run tested Unsloth GGUFs at Q8_0 (29 GB), Q4_K_M (17 GB), UD-Q2_K_XL (10.7 GB) and UD-IQ1_S (6.2 GB) against the 55 GB BF16 on GPQA Diamond, IFBench and Terminal-Bench 2.1, about $3,000 of rented GPU time. Q4_K_M matched BF16 on Terminal-Bench, and on a 24 GB card it leaves room for roughly 64k tokens of F16 KV cache. The 2-bit dropped noticeably on the coding tasks and wrote about a quarter more tokens to solve the same ones. The 1-bit scored around random guessing, and did worse at higher effort because it reasoned until the budget ran out and returned nothing.

The other write-up is about the default. Qwen ships reasoning_effort at xhigh. On a laptop, one SVG prompt took 21 minutes and 22,276 reasoning tokens; the same prompt with reasoning off took 137 seconds. LM Studio's default 8,192 token context got eaten by thinking on trivial prompts.

Hosted, it bites differently. Cerebras lists the model at about 1,850 tokens a second, $0.99 in and $1.49 out per million, reasoning on by default at high, which maps to Qwen's xhigh, and cached tokens count toward its total tokens per minute limit.

On 24 GB: Q4_K_M, set the context size yourself, start at low or medium and raise effort per task.