imatrix A/B — all-MiniLM-L6-v2 (cstr/all-MiniLM-L6-v2-GGUF), cos vs full-precision gold, n=5, calib=10 q8_0 imatrix=0 cos_vs_f16=0.999622 24.7MB -> (A/B only) q4_k imatrix=0 cos_vs_f16=0.973741 19.4MB -> (A/B only) q4_k imatrix=1 cos_vs_f16=0.974477 19.4MB -> all-MiniLM-L6-v2-q4_k-imatrix.gguf iq4_xs imatrix=1 cos_vs_f16=0.978770 19.3MB -> all-MiniLM-L6-v2-iq4_xs.gguf