Qwen3.5-27B dense Unsloth KLD table for dense versus MoE elasticity
Parent: Mac local LLMs: Quantization evaluation · Published reference · snapshot 2026-10-05
↓ Facts as markdownall context files
The Unsloth Qwen3.5 GGUF benchmarks page links the Qwen3.5-27B GGUF repo but its benchmark charts and tables are for 35B-A3B and 122B-A10B; the page states new benchmarks for 122B-A10B and 35B-A3B.
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
- The Unsloth Qwen3.5 GGUF benchmarks page links the Qwen3.5-27B GGUF repo but its benchmark charts and tables are for 35B-A3B and 122B-A10B; the page states new benchmarks for 122B-A10B and 35B-A3B. [source]
- localbench's Qwen 3.6 27B post puts Q8_0 KL at 0.075, against 0.069 for Qwen 3.6 35B-A3B and 0.120 for Qwen 3.5 27B, all on its 250K-token mixed set. [source]
- On that set the dense Qwen 3.5 27B Q8_0 floor is about 1.7 times the Qwen 3.6 27B floor, so a dense model can carry a higher irreducible KLD at near-lossless size. [source]
- 5 Mar 2026: Unsloth reissued Qwen3.5 35B, 27B, 122B and 397B with a new quantization algorithm and new imatrix data and asked users to redownload; earlier 27B numbers describe superseded files. [source]
- Unsloth's tensor findings (do not quantize ssm_out; ffn_down_exps is sensitive) are written for the MoE hybrid; a dense 27B has no expert tensors, so those rules do not transfer directly. [source]
- localbench KL and Unsloth KLD use different datasets and estimators, so the 0.120 Q8_0 value cannot be placed on the Unsloth GB axis. [source]
- Unsloth's own guidance is to use 27B when accuracy matters and 35B-A3B when speed matters, which is a quality claim without a published 27B KLD row. [source]
- Unsloth's Experiments repo may hold 27B KLD files, but the page text does not list 27B there. [source]
- The dense-versus-MoE elasticity stays unmeasured; the nearest data is localbench per-quant KL for Qwen 3.6 27B (87 quants), whose tables are images. [source]
- The Unsloth Qwen3.5 GGUF benchmark page published no Qwen3.5-27B KLD table as of the cached copy (2026-10-04). [source]
- localbench Q8_0 KL: Qwen 3.5 27B 0.120, Qwen 3.6 27B 0.075, Qwen 3.6 35B-A3B 0.069. [source]
- Unsloth's 5 Mar 2026 update replaced the Qwen3.5 27B GGUFs with a new algorithm and imatrix. [source]
- The dense-versus-MoE elasticity question has no published table to answer it. [source]
Corrections and disagreements
- CONTRADICTS: log-log-size-to-kld-elasticity-of-k-quants-acros.md Open questions line "Qwen3.5-27B has an Unsloth table not read here": the cached 2026-10-04 page has no 27B KLD rows to read. [source]
Children
- No children recorded.