Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses

I test Unsloth GGUFs of Qwen3.8 27B (Q4_K_M, UD-Q2_K_XL, UD-IQ1_S) with llama.cpp on GPQA Diamond, IFBench, Terminal-Bench 2.1. Q4_K_M matches BF16 abd fits an RTX 4090.