Importance-matrix presence as a second axis in equal-size KLD comparisons
Parent: Mac local LLMs: Quantization evaluation · Published reference · snapshot 2026-10-05
↓ Facts as markdownall context files
The covered fact is the llama.cpp scoreboard; the one thing absent is that vendor tables mix the two axes silently, shown below.
These notes link each claim to its source. Some citations identify archived research reports that are no longer hosted on this site. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.
Facts
- The covered fact is the llama.cpp scoreboard; the one thing absent is that vendor tables mix the two axes silently, shown below. [source]
- Unsloth's Qwen3.5-35B-A3B table lists Quantizer, Quant Level, Disk Space, PPL, KLD 99.9% and Mean KLD, with no imatrix column, so rows from different vendors cannot be split into recipe effect and imatrix effect from the table alone. [source]
- The same page says imatrix "definitely" lowers KLD and PPL at 5-10% slower inference, then attributes the 5-10% slowdown to I-quants (iq3_xxs, iq2_s) in its speed table; the two statements disagree about the cause. [source]
- Unsloth (imatrix helps most at low bits, works on all widths) versus the llama.cpp README (no consistent gain from more imatrix tokens, and the benefit measured on Wikitext). Not averaged; the corpora differ. [source]
- A same-recipe, same-size pair (one GGUF type, with and without imatrix, imatrix on a held-out corpus versus in-domain) scored on Qwen3.5-class MoE: none found. [source]
- Already covered by size-matched-kld-reporting-for-local-quants.md (imatrix as a second axis at fixed size, with the q4_K_S and q2_K numbers) and unsloth-dynamic-gguf-methodology.md; nothing in them is contradicted. [source]
- Unsloth's Qwen3.5-35B-A3B full-benchmark table has no per-row imatrix status column. [source]
- The Unsloth page attributes a 5-10% inference slowdown to imatrix in a bullet and to I-quant types in its speed table. [source]
Children
- No children recorded.