Qwen3.8 27B Uncensored GGUF

Uncensored Qwen3.8-27B, published as GGUF quantizations with the multi-token prediction (MTP) head retained and verified.

Refusal behaviour has been substantially reduced, not eliminated. See Measured behaviour for the numbers. Capabilities, training data, and architecture are otherwise unchanged.

Method

• Refusal directions removed with Heretic, which co-minimizes refusal count against KL divergence from the base model. No hand-written refusal-removal code, no fine-tuning, no additional training data.
• Abliteration runs at bf16 (no 4-bit quantization); the resulting LoRA is merged into the bf16 base, so the published weights are not a quantized round trip.
• mtp.* tensors are copied verbatim from the base checkpoint after merging. Abliteration never touches them — it modifies attn.o_proj and mlp.down_proj in the main stack.
• The draft head was trained against the unmodified model, so acceptance rate may fall slightly. Speculative decoding verifies every token against the target, so output quality is unaffected.
• imatrix is computed directly from the f16, not from an intermediate quantization, so calibration sees the real weights.

What's here

The draft head stays at Q8_0 in every configuration. It is small relative to the target, and quantizing it harder costs draft acceptance rate for almost no disk saving.

Overview

Files

Perplexity

Measured on this build, every quant in one session against the same f16 baseline, so the rows are comparable to each other.

Read the error bars before reading the ordering. Every row except IQ2_M sits inside a span of 0.026 against a standard error of roughly 0.25, so those quants are not separable from the f16 or from each other, and their ordering here is noise. Do not conclude that Q8_0 is worse than Q5_K_M. The only difference this measurement actually resolves is IQ2_M, which is about 2.8 standard errors above the baseline.

The noMTP-* twins are not listed because they measure identically to their fused counterparts. The MTP block is inert during a normal forward pass, which was confirmed here: fused and noMTP IQ2_M both return 7.8581, and fused and noMTP f16 both return 7.1557.

Corpus is Salesforce/wikitext, wikitext-2-raw-v1, test-00000-of-00001.parquet, text column joined with \n. This is the same file used for the importance matrix, described in full below.

llama-perplexity -m Qwen3.8-27B-Uncensored-IQ2_M.gguf \
  -f calibration.txt -ngl 99 --chunks 20

Perplexity detects gross quantization damage and nothing else. It does not measure reasoning, code, multilingual ability, or refusal behaviour. See Caveats that matter.

Importance matrix

Qwen3.8-27B-Uncensored-imatrix.dat is the importance matrix every quantization in this repo was built with. All twelve of the IQ2_M, IQ4_XS, Q4_K_M, Q5_K_M, Q6_K and Q8_0 files, fused and noMTP alike, record it in their own metadata. The standalone draft-Q8_0 head and the vision-f16 projector do not, because neither was built with one.

It is published so the files here can be reproduced, and so you can build quants that are not in this repo.

Two things to know before you use it:

• It is GGUF-format despite the .dat extension (general.type = imatrix). llama.cpp builds predating GGUF imatrix support will not load it.
• It contains no entries for blk.64, the MTP block. llama-imatrix never activates the draft head during a normal forward pass, so no activations are collected for it. This matters below.

Provenance is checkable rather than asserted: re-quantizing the f16 to Q4_K_M with this file reproduces the published Qwen3.8-27B-Uncensored-Q4_K_M.gguf to byte-identical tensors across all 866 tensors.

Building other low-bit quants yourself

The floor here is IQ2_M at 10.6 GB. If you want something smaller, or a type that is not published, the imatrix above lets you build it.

You will need the f16 GGUF, which is not published here because it is 54 GB. Build it from the bf16 weights, which are public:

hf download JonathanColetti/Qwen3.8-27B-Uncensored --local-dir Qwen3.8-27B-Uncensored
python convert_hf_to_gguf.py Qwen3.8-27B-Uncensored \
  --outfile Qwen3.8-27B-Uncensored-f16.gguf --outtype f16

Add --no-mtp to that command for the noMTP variant. The MTP shard is already grafted into the bf16 repo, so nothing needs restoring first.

The MTP block must be pinned. Because the imatrix has no blk.64 entries, and because IQ3_XXS, IQ2_XXS, IQ2_S and IQ2_M require per-tensor importance data, a fused low-bit build without a pin does not merely degrade the draft head. llama-quantize refuses to run at all. Pinning blk.64 to q8_0 sidesteps the requirement and keeps the draft head intact:

llama-quantize \
  --imatrix Qwen3.8-27B-Uncensored-imatrix.dat \
  --tensor-type 'blk\.64\.=q8_0' \
  --token-embedding-type q4_K \
  Qwen3.8-27B-Uncensored-f16.gguf Qwen3.8-27B-Uncensored-IQ2_XXS.gguf IQ2_XXS

--token-embedding-type q4_K is the largest size lever on this model. llama.cpp force-bumps token_embd to Q5_K on every IQ2/IQ1 ftype, and at 248320 vocab that is roughly 8 to 10% of parameters. Do not go below q4_K. Omit the --tensor-type pin for noMTP-* builds, since there is no block to pin.

Verify afterwards that the block survived, rather than assuming it did:

python quantize.py inspect Qwen3.8-27B-Uncensored-IQ2_XXS.gguf   # expect 65/65, has_mtp: true

Third party quants

A third party publishes a range of mixed-precision variants derived from this model:

• zerodigest/Qwen3.8-27B-Uncensored-YMQ-MTP-GGUF

Not produced by or affiliated with this repo. It is not built by me, I have not verified its files or its published numbers, and any metrics quoted there were not measured on the harness used here, so they are not comparable to the perplexity figures in this card. Linked because people ask for sizes I do not ship, not as an endorsement.

Usage

llama-server -m Qwen3.8-27B-Uncensored-Q4_K_M.gguf \
  --spec-type draft-mtp --spec-draft-n-max 2 \
  -ngl 99 -c 8192

Target plus explicit draft:

llama-server -m Qwen3.8-27B-Uncensored-noMTP-Q4_K_M.gguf \
  --spec-type draft-mtp \
  --model-draft Qwen3.8-27B-Uncensored-draft-Q8_0.gguf \
  -ngl 99 -c 8192

--spec-draft-n-max defaults to 3. Throughput depends on your hardware, so sweep it — measurements across draft lengths are in qwen3.8-spec-decode-bench.

Verification

Each artifact was checked post-quantization for MTP tensor survival rather than inferred from the conversion flag:

python quantize.py inspect Qwen3.8-27B-Uncensored-Q4_K_M.gguf

This reports metadata keys, declared block_count, and blocks actually present. A fused file whose present-block count does not exceed its declared count did not retain the MTP block.

Measured behaviour

Benchmarked against the unmodified base model on identical settings. The delta is the figure that matters: it isolates what the weight edit cost.

0-shot via lm-evaluation-harness, bf16, both models scored in the same session. Every delta is within or close to the reported standard error (MMLU +/- 0.30, ARC +/- 1.44, HellaSwag +/- 0.38, Winogrande +/- 1.21), so none is clearly separable from run-to-run noise.

These are 0-shot and are not comparable to Qwen's published scores, which use few-shot prompting. They are directly comparable to each other, which is the point. Note also that ARC-Challenge is low for a model at this MMLU — the base scores 58.9 under the same settings, so that is format sensitivity in a reasoning-tuned model, not abliteration damage.

What the benchmarks do not cover: no generative evaluation (GSM8K, HumanEval), no math or code, no multilingual, and the harness loads the text stack only — nothing here measures the vision tower or MTP speculative decoding.

Search: 200 Heretic trials, 23 non-dominated points. The published model is the marked row.

How to read these

Refusal rate is the count of refusals over 100 held-out prompts from mlabonne/harmful_behaviors (test split) — explicitly harmful requests, not benign ones. So this number is not an over-refusal rate: it does not tell you how often the model declines legitimate work. It tells you how much of the original safety behaviour on harmful requests remains.

KL divergence is measured against the unmodified base model over first-token distributions, and is the optimizer's proxy for "how much did we damage the model". Lower is closer to base. It is a proxy, not a capability measurement — a low KL does not certify that reasoning or coding ability survived, and nothing here does certify that.

The two trade off against each other. Heretic searches a Pareto front between them; the published point is one choice on that front, not a global optimum.

Caveats that matter

• Refusals were measured in non-thinking mode. This model's chat template opens a <think> block, so the evaluation closes it explicitly to score answers rather than reasoning traces. With thinking enabled the refusal rate may differ, in either direction.
• The measurement is 100 prompts from one dataset. It generalizes to that distribution of harmful requests and no further. Refusal behaviour on other topics is uncharacterized.
• Perplexity is wikitext-2 only (see the Files table). It detects gross quantization damage. It does not detect capability loss on reasoning, code, or multilingual work.
• Quantization compounds everything above. The measurements were taken on the bf16 merge; the files you download are quantized.

Requirements

MTP speculative decoding landed in llama.cpp PR #22673. Builds older than that will load these files and silently ignore the MTP tensors.

Limitations

• Refusals are reduced, not eliminated, and not redirected. This model attempts many requests the original declines, but a meaningful fraction still get refused — see Measured behaviour.
• Behaviour near the old refusal boundary is less stable than the base model.
• Lower quants compound that, and IQ2_M compounds it most. Evaluate behaviour on Q6_K or Q8_0, not on IQ2_M or IQ4_XS. Nothing here measures the refusal boundary at 2-bit.
• Capability benchmarks show a 0.5-point mean drop vs base across MMLU, ARC-Challenge, HellaSwag and Winogrande. See Measured behaviour. No generative, math, code, or multilingual evaluation was run.

Intended use

Local inference. Not intended for deployment to third parties without your own safety layer.

License

Apache 2.0, inherited from Qwen/Qwen3.8-27B. The base model's license and acceptable use policy still apply to your use of this derivative.

Speculative decoding, measured on this model

IQ2_M

Measured on this build: NVIDIA H200 NVL, 256 generated tokens, median of 3 repetitions, n_max swept 1 to 3. The table above dates from the original release and its hardware is not recorded, so compare the ratios rather than the absolute rates.

The MTP head survives 2-bit quantization because it is pinned to q8_0 rather than quantized with the rest of the stack, so speculative decoding still pays here.

Pairing noMTP-IQ2_M with the published draft-Q8_0 reaches 97.7 tok/s on the prose prompt at n_max 2, or 1.30x, which beats the fused file on that prompt. The reason is that the fused MTP head shares the low-bit token_embd (q4_K) and output (Q5_K) tensors with the main model, while the standalone draft carries its own Q8_0 copies. The split setup needs 13.3 GB of weights against 10.6 GB for the fused file, so it is the better option only if you have the VRAM to spare.