Log-log size-to-KLD elasticity of K-quants across model families
Parent: Mac local LLMs: Quantization evaluation · Published reference · snapshot 2026-10-05
↓ Facts as markdownall context files
On Qwen3.5-35B-A3B, Unsloth's own rows give a least-squares fit over 13 quants (IQ2_XXS 9.09 GB to Q8_K_XL 36.04 GB) of e of about -3.3; bartowski's 8 rows give about -3.5; segment slopes between adjacent bartowski K-quants are -4.2 (Q3_K_M to Q4_K_M) and -3.5 (Q4_K_M to Q5_K_M); AesSedai's four ...
These notes link each claim to its source. A source may be a research report hosted on this site rather than the primary document. 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
- On Qwen3.5-35B-A3B, Unsloth's own rows give a least-squares fit over 13 quants (IQ2_XXS 9.09 GB to Q8_K_XL 36.04 GB) of e of about -3.3; bartowski's 8 rows give about -3.5; segment slopes between adjacent bartowski K-quants are -4.2 (Q3_K_M to Q4_K_M) and -3.5 (Q4_K_M to Q5_K_M); AesSedai's four rows give -3.7, -3.9 and -3.0. [source]
- Individual Unsloth segments scatter widely (from -0.4 for Q5_K_XL to Q6_K_S to -14 for Q4_K_M to Q4_K_L) because each recipe moves different tensors, so a single pair is not an elasticity. [source]
- Computed from the same table: Q3_K_M to Q4_K_M to Q5_K_XL on Unsloth rows give -3.0 then -4.5. [source]
- Same-label spread: Q4_K_M is 18.49 GB with mean KLD 0.0192 (Unsloth), 19.77 GB with 0.0182 (bartowski) and 20.62 GB with 0.0096 (AesSedai). AesSedai is 11.5% larger than Unsloth with half the KLD, which is a slope of -6.3 between two recipes at the same label. [source]
- At e of -3.3, a 3% size gap explains about a 10% KLD gap on this MoE, against about 20% on LLaMA 3 8B. [source]
- 2024: LLaMA 3 8B scoreboard (existing dossier). 2026-03: Unsloth publishes 150+ KLD runs on Qwen3.5 (existing dossier). This file is the first to put them on one log-log axis. [source]
- The table mixes imatrix and recipe differences and has no imatrix column, so the fitted slope includes both. [source]
- IQ and K families lie on different lines; the Unsloth fit pools them. [source]
- 99.9th-percentile KLD in the same table orders quants differently from mean KLD in places (for example MXFP4_MOE 18.17 GB has mean 0.0272 and KLD 99.9 of 0.7789 against Q4_K_M 18.49 GB at 0.0192 and 0.5478), so tail elasticity needs its own fit. [source]
- A dense 27B-class model (Qwen3.5-27B has an Unsloth table not read here) to separate dense from MoE. [source]
- Elasticity for Gemma 4 31B, whose ranked-KLD thread could not be fetched (reddit unsupported by the fetcher). [source]
- Fitted log-log slope of mean KLD against GB is about -3.3 for Unsloth's 13 Qwen3.5-35B-A3B quants and about -3.5 for bartowski's 8. [source]
- bartowski's Qwen3.5-35B-A3B Q3_K_M (14.95 GB, 0.0585), Q4_K_M (19.77 GB, 0.0182) and Q5_K_M (23.11 GB, 0.0106) give segment slopes of -4.2 and -3.5. [source]
- AesSedai's IQ4_XS (16.4 GB, 0.0235), Q4_K_M (20.62 GB, 0.0096) and Q5_K_M (24.45 GB, 0.0058) give segment slopes of -3.9 and -3.0. [source]
- Q4_K_M on Qwen3.5-35B-A3B ranges from 18.49 GB at mean KLD 0.0192 (Unsloth) to 20.62 GB at 0.0096 (AesSedai). [source]
- The LLaMA 3 8B slope of about -7 (from q5_K_M 5.33 GiB and 0.010762 to q4_K_M 4.58 GiB and 0.031273) is not a constant across families: the Qwen3.5 MoE slope is about half. [source]
Children
- No children recorded.