Quant makers, head to head

The same model at the same quant type, uploaded by different people, is not the same file: makers use different importance-matrix data, tensor-type overrides and llama.cpp versions. Here we only compare files that we measured, at the same model and the same llama.cpp quant type.

KLD intervals: mean KLD ± 1.96 × the standard error printed by llama-perplexity (25,500 positions). Overlapping intervals are treated as a tie. Because every quant is scored on the same tokens, these per-file intervals are conservative. "i1-" (mradermacher) marks an imatrix quant of the same type.

Granite 4.2 8B · Q4_K_M (GGUF: bartowski vs mradermacher)

Lowest KLD: bartowski (0.0729). mradermacher: higher (CIs do not overlap). Accuracy: all tied within 95% intervals. Tool calls: all tied within 95% intervals.

Qwen3.5 4B · Q4_K_M (GGUF: bartowski vs mradermacher vs unsloth)

Lowest KLD: bartowski (0.0258). unsloth: higher (CIs do not overlap); mradermacher: higher (CIs do not overlap). Accuracy: all tied within 95% intervals. Tool calls: all tied within 95% intervals.

Qwen3.5 9B · IQ4_XS (GGUF: bartowski vs mradermacher)

Lowest KLD: bartowski (0.0332). mradermacher: tied (95% CIs overlap). Accuracy: all tied within 95% intervals. Tool calls: all tied within 95% intervals.

Qwen3.5 9B · Q4_K_M (GGUF: bartowski vs mradermacher vs unsloth)

Lowest KLD: bartowski (0.0219). unsloth: higher (CIs do not overlap); mradermacher: higher (CIs do not overlap). Accuracy: all tied within 95% intervals. Tool calls: all tied within 95% intervals.

Not comparable yet

These quant types were measured from only one maker, so there's no fair head-to-head yet. Different quant types (e.g. unsloth UD-Q4_K_XL vs bartowski Q4_K_M) are not compared here because they aren't the same bit recipe; see each model page for those.

Want a specific matchup? Request it. Proposed next comparisons are queued for approval; results appear here only once measured.