Qwen3.8 delivers substantial gains across coding, professional work, research, and long-horizon agentic tasks.

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
Whats here
Overview
Files
Perplexity
- Baseline against the unmodified base model
- How the baseline was measured
Importance matrix
Building other low bit quants yourself
- Third party quants
Usage
- ComfyUI
- llama cpp
Verification
Measured behaviour
- How to read these
- Caveats that matter
Requirements
Limitations
License
Speculative decoding, measured on this model
- IQ2_M

Method

Refusal directions removed with Heretic, which co minimizes refusal count against KL divergence from the base model. No handwritten refusal removal code, no finetuning, 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. Instead abliteration 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.

Whats here

Family	Files	Use when
Fused	`Qwen3.8-27B-Uncensored-<QUANT>.gguf`	One file. MTP rides inline as a built-in draft.
Target + draft	`Qwen3.8-27B-Uncensored-noMTP-<QUANT>.gguf` + `Qwen3.8-27B-Uncensored-draft-<QUANT>.gguf`	Your runtime wants an explicit `--model-draft`.
Vision	`mmproj-Qwen3.8-27B-Uncensored-F16.gguf`	Image input with a compatible vision runtime.
MTP Head for llama server	`Qwen3.8-27B-Uncensored-draft-<QUANT>.gguf`	MTP head for llama-server, specifically for the flag `--model-draft`

The draft head ships at Q8_0 and Q4_0, the same weights at two precisions. Q8_0 is the default and is what the speculative decoding tables below were measured with. Q4_0 saves 1.5 GB for tight VRAM budgets, and its acceptance rate has not been measured here.

The F16 vision projector uses the standard mmproj prefix for automatic discovery by compatible runtimes.

Overview


Base	Qwen/Qwen3.8-27B
Architecture	`Qwen3_5ForConditionalGeneration`
Layers	64
Vocab	248320
MTP layers	1
Vision	yes
Context	262144
Quants	IQ2_M, IQ4_XS, Q4_K_M, Q5_K_M, Q6_K, Q8_0
imatrix	wikitext-2 raw, 200 chunks, published
Converted with	llama.cpp `a94d563ed`

Files

File	Size	MTP	PPL (wikitext-2)
`Qwen3.8-27B-Uncensored-IQ2_M.gguf`	10.6 GB	yes	PPL = 7.8581 +/- 0.27481
`Qwen3.8-27B-Uncensored-IQ4_XS.gguf`	15.3 GB	yes	PPL = 7.1583 +/- 0.25019
`Qwen3.8-27B-Uncensored-Q4_K_M.gguf`	16.8 GB	yes	PPL = 7.1814 +/- 0.25227
`Qwen3.8-27B-Uncensored-Q5_K_M.gguf`	19.5 GB	yes	PPL = 7.1573 +/- 0.25055
`Qwen3.8-27B-Uncensored-Q6_K.gguf`	22.4 GB	yes	PPL = 7.1689 +/- 0.25149
`Qwen3.8-27B-Uncensored-Q8_0.gguf`	29.0 GB	yes	PPL = 7.1764 +/- 0.25195
`Qwen3.8-27B-Uncensored-draft-Q4_0.gguf`	1.7 GB	-	-
`Qwen3.8-27B-Uncensored-draft-Q8_0.gguf`	3.2 GB	-	-
`Qwen3.8-27B-Uncensored-noMTP-IQ2_M.gguf`	10.2 GB	no	PPL = 7.8581 +/- 0.27481
`Qwen3.8-27B-Uncensored-noMTP-IQ4_XS.gguf`	15.1 GB	no	-
`Qwen3.8-27B-Uncensored-noMTP-Q4_K_M.gguf`	16.5 GB	no	-
`Qwen3.8-27B-Uncensored-noMTP-Q5_K_M.gguf`	19.2 GB	no	-
`Qwen3.8-27B-Uncensored-noMTP-Q6_K.gguf`	22.1 GB	no	-
`Qwen3.8-27B-Uncensored-noMTP-Q8_0.gguf`	28.6 GB	no	-
`mmproj-Qwen3.8-27B-Uncensored-F16.gguf`	0.9 GB	-	-
`Qwen3.8-27B-Uncensored-imatrix.dat`	13.6 MB	-	-

Perplexity

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

File	PPL (wikitext-2)	vs f16
`Qwen3.8-27B-Uncensored-f16.gguf` (baseline, not shipped)	7.1557 +/- 0.25104
`Qwen3.8-27B-Uncensored-Q5_K_M.gguf`	7.1573 +/- 0.25055	+0.0016
`Qwen3.8-27B-Uncensored-IQ4_XS.gguf`	7.1583 +/- 0.25019	+0.0026
`Qwen3.8-27B-Uncensored-Q6_K.gguf`	7.1689 +/- 0.25149	+0.0132
`Qwen3.8-27B-Uncensored-Q8_0.gguf`	7.1764 +/- 0.25195	+0.0207
`Qwen3.8-27B-Uncensored-Q4_K_M.gguf`	7.1814 +/- 0.25227	+0.0257
`Qwen3.8-27B-Uncensored-IQ2_M.gguf`	7.8581 +/- 0.27481	+0.7024

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

Baseline against the unmodified base model

Every row above is this model, so that table shows what quantization costs and nothing more. It does not show what the abliteration costs. That needs the unmodified base model run on the same harness, which is what follows.

Model	File	PPL (wikitext-2)	vs base bf16
Qwen/Qwen3.8-27B	`Qwen3.8-27B-BF16.gguf`	6.5129 +/- 0.04209	baseline
Qwen/Qwen3.8-27B	`Qwen3.8-27B-Q8_0.gguf`	6.5122 +/- 0.04208	-0.0007
This model	`Qwen3.8-27B-Uncensored-BF16.gguf` (not shipped)	6.5563 +/- 0.04248	+0.0434

Base files come from ggml-org/Qwen3.8-27B-GGUF. The bf16 GGUF of this model was converted locally from the published bf16 weights and is not shipped here, for the same reason the f16 is not.

The weight edit costs 0.0434 perplexity, about 0.7% relative. Measured as a paired difference over the same tokens rather than as two independent runs, the gap is 0.0437 +/- 0.0163, so it is a real cost rather than noise. The base Q8_0 row is the control: it lands 0.0007 from the base bf16, well below the gap being reported.

Distribution overlap was measured directly rather than inferred from the perplexity gap:


Mean KL divergence, base bf16 against this model at bf16	0.0141 +/- 0.0025
Maximum KL divergence	19.956
Same top token	96.72%
Tokens compared	20480

This is a different measurement from the first token KL under Measured behaviour, which is the figure Heretic optimizes against. This one runs over ordinary corpus text and covers every position, not just the first.

How the baseline was measured

One session, one machine, one binary, the same corpus file and the same settings for every row, with each model fully offloaded:

llama-perplexity -m Qwen3.8-27B-BF16.gguf \
  -f calibration.txt -c 2048 -b 2048 -ngl 99


Corpus	`calibration.txt`, byte identical to the file under Importance matrix, md5 `d998c24b049cf7c009dbf2672da70b5a`
Coverage	whole file, 145 chunks of 2048 tokens
Built with	llama.cpp `a94d563ed`
Hardware	NVIDIA H200 NVL, full offload at `-ngl 99`

These numbers do not line up with the table above. That table stops at 20 chunks and this one reads the whole file, so the two cover different amounts of text and cannot be compared directly. Rows within each table are comparable to each other. The whole file run carries a standard error of roughly 0.042 against 0.25 for the 20 chunk run, which is why a gap this small is worth reporting at all.

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 mmproj-Qwen3.8-27B-Uncensored-F16.gguf 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.


Corpus	`Salesforce/wikitext`, `wikitext-2-raw-v1`, `test-00000-of-00001.parquet`
Assembly	`text` column joined with `\n`, giving 1,292,013 bytes, md5 `d998c24b049cf7c009dbf2672da70b5a`
Chunks	200 x 512 tokens
Computed from	the f16 GGUF, not an intermediate quantization
Built with	llama.cpp `a94d563ed`

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: requantizing 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

2 bit warning. IQ2_M is the most degraded file in this repo, and anything you build below it will be worse. Expect the loss to land hardest on the thing this model is used for. Behaviour near the old refusal boundary is already its least stable property (see Caveats that matter), and 2 bit compounds exactly that. Perplexity will tell you the model is not broken. It will not tell you the refusal boundary still behaves the way it does at Q6_K, and nothing in this repo measures that at 2 bit.

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

ComfyUI

Tested versions:

ComfyUI commit 0a33ed6c28f926d14536235771c222f9e6d1026b
ComfyUI QwenVL commit e79582111e5835787574bcc17adc4bf4cf3a07f4
llama cpp python version 0.3.48 at commit c57b174711166b889f06a35a30af95a75db24705

# Install ComfyUI and the QwenVL nodes
git clone https://github.com/comfyanonymous/ComfyUI.git
git -C ComfyUI checkout 0a33ed6c28f926d14536235771c222f9e6d1026b
git clone https://github.com/1038lab/ComfyUI-QwenVL.git ComfyUI/custom_nodes/ComfyUI-QwenVL
git -C ComfyUI/custom_nodes/ComfyUI-QwenVL checkout e79582111e5835787574bcc17adc4bf4cf3a07f4

# Install dependencies
uv pip install -r ComfyUI/requirements.txt
uv pip install -r ComfyUI/custom_nodes/ComfyUI-QwenVL/requirements.txt
CMAKE_ARGS="-DGGML_CUDA=on -DLLAMA_CPP_PYTHON_VISION=on" \
  uv pip install \
  "llama-cpp-python @ git+https://github.com/JamePeng/llama-cpp-python.git@c57b174711166b889f06a35a30af95a75db24705"

# Download the language model and F16 vision projector
hf download JonathanColetti/Qwen3.8-27B-Uncensored-GGUF \
  Qwen3.8-27B-Uncensored-Q4_K_M.gguf \
  mmproj-Qwen3.8-27B-Uncensored-F16.gguf \
  --local-dir \
  ComfyUI/models/LLM/GGUF/JonathanColetti/Qwen3.8-27B-Uncensored-GGUF

# Start ComfyUI
cd ComfyUI
python main.py --listen 127.0.0.1 --port 8188

Add the model to ComfyUI/custom_nodes/ComfyUI-QwenVL/custom_models.json:

{
  "hf_models": {},
  "gguf_models": {
    "Qwen3.8 27B Uncensored Q4 K M": {
      "author": "JonathanColetti",
      "repo_name": "Qwen3.8-27B-Uncensored-GGUF",
      "repo_id": "JonathanColetti/Qwen3.8-27B-Uncensored-GGUF",
      "model_files": [
        "Qwen3.8-27B-Uncensored-Q4_K_M.gguf"
      ]
    }
  }
}

The projector is discovered automatically because its filename starts with mmproj.

llama cpp

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.

File	MTP	blocks
`Qwen3.8-27B-Uncensored-f16.gguf`	True	⁶⁵⁄₆₅
`Qwen3.8-27B-Uncensored-noMTP-f16.gguf`	False	⁶⁴⁄₆₄
`Qwen3.8-27B-Uncensored-IQ2_M.gguf`	True	⁶⁵⁄₆₅
`Qwen3.8-27B-Uncensored-noMTP-IQ2_M.gguf`	False	⁶⁴⁄₆₄
`Qwen3.8-27B-Uncensored-IQ4_XS.gguf`	True	⁶⁵⁄₆₅
`Qwen3.8-27B-Uncensored-noMTP-IQ4_XS.gguf`	False	⁶⁴⁄₆₄
`Qwen3.8-27B-Uncensored-Q4_K_M.gguf`	True	⁶⁵⁄₆₅
`Qwen3.8-27B-Uncensored-noMTP-Q4_K_M.gguf`	False	⁶⁴⁄₆₄
`Qwen3.8-27B-Uncensored-Q5_K_M.gguf`	True	⁶⁵⁄₆₅
`Qwen3.8-27B-Uncensored-noMTP-Q5_K_M.gguf`	False	⁶⁴⁄₆₄
`Qwen3.8-27B-Uncensored-Q6_K.gguf`	True	⁶⁵⁄₆₅
`Qwen3.8-27B-Uncensored-noMTP-Q6_K.gguf`	False	⁶⁴⁄₆₄
`Qwen3.8-27B-Uncensored-Q8_0.gguf`	True	⁶⁵⁄₆₅
`Qwen3.8-27B-Uncensored-noMTP-Q8_0.gguf`	False	⁶⁴⁄₆₄

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.

Task	Base	Uncensored	Δ
MMLU	83.4	83.3	-0.2
ARC-Challenge	58.9	57.7	-1.2
HellaSwag	82.8	82.9	+0.1
Winogrande	76.1	75.3	-0.8
Mean			-0.5

zero 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.

Measurement	Base model	This model
Refusals (100 held-out harmful prompts)	⁹⁸⁄₁₀₀	¹²⁄₁₀₀
KL divergence vs base (first-token)	0	0.1191

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

refusals	KL divergence
¹²⁄₁₀₀	0.1191	<= published
¹³⁄₁₀₀	0.1052
¹⁹⁄₁₀₀	0.0722
²³⁄₁₀₀	0.0635
²⁶⁄₁₀₀	0.0507
²⁷⁄₁₀₀	0.0410
³⁵⁄₁₀₀	0.0406
³⁶⁄₁₀₀	0.0387
⁴¹⁄₁₀₀	0.0366
⁴⁴⁄₁₀₀	0.0352
⁴⁶⁄₁₀₀	0.0334
⁴⁸⁄₁₀₀	0.0331
⁵¹⁄₁₀₀	0.0321
⁵²⁄₁₀₀	0.0294
⁶⁰⁄₁₀₀	0.0290
⁷⁶⁄₁₀₀	0.0280
⁷⁷⁄₁₀₀	0.0247
⁸³⁄₁₀₀	0.0204
⁸⁶⁄₁₀₀	0.0193
⁹¹⁄₁₀₀	0.0170
⁹⁶⁄₁₀₀	0.0146
⁹⁷⁄₁₀₀	0.0044
⁹⁸⁄₁₀₀	0.0004

How to read these

Refusal rate is the count of refusals over 100 heldout 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 optimizers 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 models 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.

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

prompt	spec_type	n_max	tok/s	vs baseline
prose	none	-	74.8	1.00x
prose	draft-mtp	1	89.0	1.19x
prose	draft-mtp	2	85.7	1.15x
prose	draft-mtp	3	72.0	0.96x
prose	draft-mtp	4	71.1	0.95x
prose	draft-mtp	5	62.9	0.84x
prose	draft-mtp	6	53.8	0.72x
prose	draft-mtp	7	49.9	0.67x
prose	draft-mtp	8	59.9	0.80x
code	none	-	74.7	1.00x
code	draft-mtp	1	95.4	1.28x
code	draft-mtp	2	92.9	1.24x
code	draft-mtp	3	82.6	1.11x
code	draft-mtp	4	74.9	1.00x
code	draft-mtp	5	67.4	0.90x
code	draft-mtp	6	59.4	0.80x
code	draft-mtp	7	55.6	0.74x
code	draft-mtp	8	70.9	0.95x
chat	none	-	74.7	1.00x
chat	draft-mtp	1	90.6	1.21x
chat	draft-mtp	2	84.2	1.13x
chat	draft-mtp	3	76.1	1.02x
chat	draft-mtp	4	70.4	0.94x
chat	draft-mtp	5	64.3	0.86x
chat	draft-mtp	6	55.2	0.74x
chat	draft-mtp	7	50.2	0.67x
chat	draft-mtp	8	54.1	0.72x

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.

prompt	spec_type	n_max	tok/s	vs baseline
prose	none	-	75.4	1.00x
prose	draft-mtp	1	85.2	1.13x
prose	draft-mtp	2	83.2	1.10x
prose	draft-mtp	3	77.4	1.03x
code	none	-	75.5	1.00x
code	draft-mtp	1	95.8	1.27x
code	draft-mtp	2	99.8	1.32x
code	draft-mtp	3	96.0	1.27x
chat	none	-	75.3	1.00x
chat	draft-mtp	1	87.4	1.16x
chat	draft-mtp	2	81.3	1.08x
chat	draft-mtp	3	78.4	1.04x

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.