Archival mirror: the five unsloth/Qwen3.8-27B-GGUF files our published comparison cites, at revision f1bfb127c64f7072bdd2cad55f258b9c8b2910fe

This repository exists so that a published measurement stays reproducible. It is a byte-identical mirror of five files from unsloth/Qwen3.8-27B-GGUF at one pinned revision — nothing requantized, re-exported, or edited. Use the upstream repository for anything current; it has the full quant ladder and this has only the five files a published table depends on.

Why this exists, stated precisely

At the time of writing, revision f1bfb127c64f7072bdd2cad55f258b9c8b2910fe still resolves upstream and is in fact current HEAD. So unlike our NVFP4 archival mirror, this one is precautionary rather than a recovery.

The precaution is not hypothetical. The sibling repository unsloth/Qwen3.8-27B-NVFP4 was super-squashed on 2026-08-15, and the revision our published comparison cited stopped resolving — the Hub now answers {"error":"Invalid rev id"} for it. Our local snapshot turned out to be the only verifiable copy. Squashing history is a routine maintainer action and entirely upstream's call; the lesson is ours: a Hugging Face revision id is not a durable citation, so anything a published number depends on gets mirrored while it still exists rather than after it does not.

Mirroring is also nearly free, which is what makes it a sensible default rather than a special measure. Hugging Face Xet storage is content-addressed globally, so re-uploading bytes that already exist on the Hub transfers almost nothing: our 22.5 GB NVFP4 mirror moved about 205 MB on the wire.

Files and digests

Every file was verified locally against the published lfs_sha256 of revision f1bfb127 before upload, and re-verified from the Hub's own tree API afterwards.

Total 126.8 GB. Upstream paths are preserved, so the two-part BF16 set still loads as BF16/Qwen3.8-27B-BF16-00001-of-00002.gguf.

What we measured with these files

They are the GGUF rows of a teacher-forced distribution-fidelity comparison of Qwen3.8-27B quantizations: KL(BF16 reference ‖ candidate) over a held-out suite of 512 contexts and 1,048,064 scored positions, full 248,320-entry vocabulary, every candidate read at the post-final-norm state and projected through one shared BF16 LM head, so no candidate's own head quantization is counted.

• Q8_0 — 0.001087 mean KLD, the fidelity leader on that suite
• Q6_K — 0.002035
• UD-Q5_K_XL — 0.004444
• the two-part BF16 set is not a quantization row at all: it measures the cross-engine floor. Captured through llama.cpp and compared against the same weights served by vLLM, it answers "what do two engines disagree by on identical unquantized weights?" The answer is 0.000507 mean KLD, 99.07 % top-1 agreement, and every GGUF row above contains that term while the vLLM-side rows do not. Without this file pair, none of the GGUF numbers could be read net of engine numerics — which is exactly why it is mirrored alongside them.

Full method, intervals and caveats are in the receipts/ tree of the qwen38-27b project: receipts/cross-engine-comparator.json, receipts/gguf-report-q8_0.json, receipts/gguf-report-q6_k.json, receipts/gguf-report-q5_k_xl.json, receipts/gguf-report-engine-floor.json.

Attribution and licence

Quantization and packaging by unsloth; base model Qwen/Qwen3.8-27B. Licence apache-2.0, inherited unchanged from upstream, as are the base_model and unsloth tags. All credit for these artifacts belongs upstream; this repository contributes nothing but durability.