Qwen 3.6 27B GGUF quality benchmark per-quant KL (localbench, 87 quants)
Parent: Mac local LLMs: Quantization evaluation · Published reference · snapshot 2026-10-05
↓ Facts as markdownall context files
Uploaders and counts: unsloth 21, lmstudio-community 3, Jackrong 11, bartowski 26, mradermacher (i1) 23, ubergarm 2 (ik_llama.cpp), ggml-org 1.
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
- Uploaders and counts: unsloth 21, lmstudio-community 3, Jackrong 11, bartowski 26, mradermacher (i1) 23, ubergarm 2 (ik_llama.cpp), ggml-org 1. [source]
- Method is the shared localbench one: about 250,000 tokens in six categories, top-40 KL on prompt tokens through the model's chat template. [source]
- The frontier table lists, for each size, the lowest-KL quant; a quant absent from it is dominated by a smaller file with lower KL. [source]
- 25 Apr 2026: post published by oobabooga, after the 7 Apr Gemma 4 31B post. [source]
- Long documents carry most of the loss: UD-Q8_K_XL scores KL 0.001 on coding but 0.373 on long documents. [source]
- Tool calling is the second-worst category at every size. [source]
- ik_llama.cpp-only types (IQ5_KS, IQ4_KS) cannot run on mainline llama.cpp, so their frontier spots do not help a stock Metal build. [source]
- The cache shows only the page text; the Pareto table and per-quant numbers sit in images, so rank order within the frontier is unknown here. [source]
- Existing Gemma 4 31B result: unsloth holds 8 of 9 frontier spots. Qwen 3.6 27B: bartowski and mradermacher tie at 12 each and unsloth has 10, so no uploader dominates. Both are in the same author's series. [source]
- The benchmark covers 87 Qwen 3.6 27B GGUF quants from 7 uploaders against the unsloth BF16 reference. [source]
- bartowski and mradermacher each hold 12 frontier positions and unsloth holds 10. [source]
- ubergarm's two ik_llama.cpp quants both reach the frontier: IQ5_KS (19.9 GB, KL 0.128) and IQ4_KS (15.8 GB, KL 0.209). [source]
- Jackrong has no frontier quants out of 11 and lmstudio-community has none out of 3. [source]
- Q8_0 has KL 0.075, close to Qwen 3.6 35B-A3B (0.069) and better than Qwen 3.5 27B (0.120). [source]
- IQ2_XXS (8.4 GB) keeps 77.5% top-1 agreement with BF16. [source]
- The post asks readers to share the URL, not its plots and tables, because the measurements are expensive and subscriptions fund more models. [source]
Children
- No children recorded.