Qwen3.8-27B quantization benchmark: Q4_K_M is free, Q2 costs 5 points on agentic coding, 1-bit collapses to random chance.

Quesma ran the model across GPQA Diamond, IFBench and Terminal-Bench 2.1 (89 agentic coding tasks) on L40S, H100 and H200 via Modal. Q4_K_M at 17 GB matched BF16 at 55 GB within a point on all three. UD-Q2_K_XL at 10.7 GB held instruction-following but dropped Terminal-Bench from about 77% to about 72%. Both 1-bit quants at 6.2 GB fell to 15-20% on GPQA Diamond, which is around random chance, and degraded further on longer reasoning (Quesma). Q4_K_M remains the default answer and this is the cleanest evidence for it I've seen this month.