Qwen3.8-27B-NVFP4-archival-9c73e2da

Archival mirror: unsloth/Qwen3.8-27B-NVFP4 at revision 9c73e2daee1d0fd494ffbd1d8753f2174a953796

This is a byte-identical archival mirror of one revision of unsloth/Qwen3.8-27B-NVFP4 that the upstream repository no longer serves. Every file here is the same bytes as that revision; nothing has been requantized, re-exported, or edited. Use the upstream repository for anything current — this mirror exists only so that a published measurement stays checkable.

Why this exists

On 2026-08-15 the upstream repository's history was super-squashed (commit 16b6615a is titled "Super-squash branch 'main' using huggingface_hub"). Squashing a model repository is routine and entirely the maintainer's call — the point of this mirror is narrower and is about us, not about unsloth: pinning a Hugging Face revision is not a durable citation. After the squash, https://huggingface.co/api/models/unsloth/Qwen3.8-27B-NVFP4/revision/9c73e2daee1d0fd494ffbd1d8753f2174a953796 answers {"error":"Invalid rev id"}, so the revision string in our published receipts stopped resolving.

We had measured this exact revision, and we cite it. So the snapshot is mirrored here, digests and all, and our receipts now cite this mirror alongside the original revision id.

Relationship to current upstream HEAD (16b6615af3548b88e2d8e382457bc705b00479cf)

The weights are byte-identical. model.safetensors and model_mtp.safetensors have the same sha256 here as at current upstream HEAD, so this is not a different quantization of the model and nothing about the served numerics differs.

Three files do differ between the mirrored revision and current HEAD:

config.json is otherwise identical field for field — that one observer entry is the only difference in the entire file, including every config_groups target, every scale strategy, and the whole ignore list.

The observer field is metadata about how the quantizer chose scales at build time, recorded by llm-compressor; it is not read at inference time, and the weights it describes are the same bytes in both revisions. So it does not change what the model computes. It is worth recording anyway, because it is the only surviving evidence of how the reviewed revision described its own construction: the revision we measured declares an imatrix-MSE observer, and the current one declares none.

The tokenizer.json difference, and why this mirror was already needed once

This mirror earned its keep within the hour. One of our own documents pinned tokenizer.json sha256 06b9509352d2af…, while our local snapshot hashed f399b3cd12fa…, and nobody could tell which side had drifted, because the revision we had cited no longer resolved. Holding the mirror settled it by measurement: 06b9509352d2af… (19,989,325 B) is current HEAD's, f399b3cd12fa… (19,989,424 B) is the reviewed revision's — the one in this repository, and it hashes byte-identically to our snapshot. Without the mirror that row would have had to stay unresolved permanently. This is the entire argument for archival mirroring in one paragraph: the point is not storage, it is that a citation stays checkable.

What actually differs between the two tokenizers is worth knowing if you use this repository, because it is inert in one API and a trap in another. A full structural walk of both 20 MB files reports exactly one differing node: truncation. This revision carries {"direction":"Right","max_length":2048,"strategy":"LongestFirst","stride":0}; HEAD carries null. The vocab (248,044 entries), merges and added_tokens are identical.

• Through transformers.AutoTokenizer, the two are equivalent — measured: a 12,002-token input tokenizes to 12,002 tokens with either file, because PreTrainedTokenizerFast sets truncation per call and defaults to off.
• Through the raw tokenizers.Tokenizer API, they are not — the file in this repository silently truncates that same input to 2,048 tokens, while HEAD's returns all 12,002.

So if you load this revision with the raw library, disable truncation explicitly or you will quietly lose everything past 2,048 tokens. None of this affects the fidelity measurement described below, which replays a pre-tokenized suite and consults no tokenizer at all.

Digests

Verified locally against the pinned snapshot before upload, and re-verified after upload.

The ten non-LFS files (config.json, model.safetensors.index.json, .gitattributes, README.md as mirrored under upstream-README.md, generation_config.json, tokenizer_config.json, preprocessor_config.json, video_preprocessor_config.json, vocab.json, chat_template.jinja) additionally match their upstream git blob ids at revision 9c73e2da, recomputed as sha1("blob <size>\0" + bytes).

The upstream card for the mirrored revision is preserved verbatim as upstream-README.md; this file replaces it as the repository card so that the archival status is the first thing a reader sees.

Attribution and licence

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

What we measured with it

unsloth/Qwen3.8-27B-NVFP4 at this revision, served under vLLM with --quantization compressed-tensors, is one candidate in a teacher-forced distribution-fidelity comparison of Qwen3.8-27B quantizations: KL(BF16 reference ‖ candidate) over a held-out suite, full 248,320-entry vocabulary, both operands read at the final norm and projected through one shared BF16 LM head so that no candidate's own head quantization is counted. Reported in the receipts/ tree of the qwen38-27b project, including receipts/nvfp4-v5-measurement.json, which records this mirror.