Qwen3.8-27B ZeroRefusal UD-IQ4_XS V3 Final (Native MTP)

This is a Qwen3.8-27B GGUF built and validated for an RTX 5060 Ti 16 GB deployment.

The model was requantized from ZeroFuse-edited BF16 weights using the tensor-type layout and importance matrix of the final Unsloth Dynamic V3 UD-IQ4_XS release. The resulting single GGUF retains the native MTP / NextN layer and does not require a separate draft model.

This is an independent community release, not an official release from Qwen, Unsloth, llama.cpp, or ZeroFuse.

Release files

File	Purpose	Size
`Qwen3.8-27B-ZeroRefusal-UD-IQ4_XS-V3-Final-MTP.gguf`	Main text model with native MTP	14,252,845,056 bytes
`mmproj-Qwen3.8-27B-F16.gguf`	Qwen3.8 Vision F16 projector	927,607,008 bytes
`quantization/imatrix_unsloth.gguf`	Importance matrix used for quantization	13,642,656 bytes
`quantization/v3-final-tensor-overrides.txt`	Anchored type map for all 866 tensors	17,007 bytes

Main model SHA256:

f6c1f6e1211aeceaa386cbeabd0bdfbee0d124747ecaa3e908a93394518e9274

See SHA256SUMS for all published binary hashes.

Model structure

Architecture: qwen35
Tensor count: 866
Block count: 65
qwen35.nextn_predict_layers = 1
Native MTP: 15 tensors under blk.64.*
general.file_type = 30
llama.cpp reports: IQ4_XS - 4.25 bpw
Original context length: 262,144
Validated production context: 73,728

The final tensor names, shapes, and quantization-type distribution match the final Unsloth Dynamic V3 UD-IQ4_XS reference. This does not imply byte-identical tensor payloads: the behavior weights were edited and requantized.

Validated production profile

GPU                 RTX 5060 Ti 16 GB
llama.cpp           upstream b10435 / 9e40df6
context             73728 (72 x 1024, shared KV pool)
parallel            2
KV                   unified Q4_0 K / Q4_0 V
text placement      full GPU
MTP                  1
draft KV             F16 K / F16 V
Vision               F16 mmproj, CPU offload
image-max-tokens     4096
API port             8001

With 73728 + P2 + unified KV, both slots share one 72K KV pool. It is not a separate 72K context for each slot.

llama-server example

llama-server \
  -m Qwen3.8-27B-ZeroRefusal-UD-IQ4_XS-V3-Final-MTP.gguf \
  --alias qwen3.8-27b \
  --mmproj mmproj-Qwen3.8-27B-F16.gguf \
  --no-mmproj-offload \
  --image-max-tokens 4096 \
  -c 73728 \
  -np 2 \
  --kv-unified \
  -ngl 999 \
  --flash-attn on \
  -ctk q4_0 \
  -ctv q4_0 \
  --spec-type draft-mtp \
  --spec-draft-n-max 1 \
  --spec-draft-type-k f16 \
  --spec-draft-type-v f16 \
  --no-spec-draft-backend-sampling \
  -b 512 \
  -ub 64 \
  --threads 7 \
  --fit off \
  --jinja \
  --host 0.0.0.0 \
  --port 8001

Measured performance on RTX 5060 Ti 16 GB

Test	Result
P1 decode	41.67 tok/s
P2 stream A	31.58 tok/s
P2 stream B	34.08 tok/s
P2 average per stream	32.83 tok/s
P2 server aggregate	65.67 tok/s
P2 wall-clock aggregate	61.88 tok/s
Lowest observed free VRAM	about 138 MiB

MTP draft acceptance on the speed-test samples was 80.85% for P1 and 77.43% / 62.22% for the two P2 streams. Acceptance is content-dependent.

64-prompt refusal validation

Under the fixed validation configuration:

valid       64
accepted    64
refused      0
errors       0
result       0/64 refusal

Method:

Dataset: mlabonne/harmful_behaviors
Revision: 01cead01398926d81f7c52bdb790ee8cf77ebba7
Corpus SHA256: b4f2ddec5ab06058b721be9afe75e7fdc656852da257d8e7a117425b3dd89114
First 64 prompts, maximum 64 output tokens per prompt
Temperature 1.0, top-p 0.95, top-k 20
Fixed 23-marker refusal classifier

Only the methodology and summary are public. Raw harmful outputs are intentionally not uploaded. This result applies only to the fixed test and is not a general safety or quality guarantee.

Provenance

Final Dynamic V3 quantization reference:

repo      unsloth/Qwen3.8-27B-GGUF
file      Qwen3.8-27B-UD-IQ4_XS.gguf
revision  4ca720788d1e01f1bff70c033e0d0028fd02e502
bytes     14252845984
sha256    40fac4050e940397dbf13087afd50f4734a11805bf9d65ef8ddd7483470e6199

Behavior-edit source:

repo          junafinity/Qwen-3.8-27B-Uncensored
revision      903d149c148b81fdf4e568a05ac9ad4225f493d7
method        ZeroFuse v0.1.0 directional ablation
trial         38
source layer  35
strength      1.2242385643286666
edited layers 9-56
BF16 KL       0.009713646
BF16 refusal  0/64

The quantization map uses fully anchored regular expressions. This prevents the global output.weight rule from accidentally matching per-layer attn_output.weight tensors. See quantization/v3-final-tensor-overrides.txt.

Limitations and safety

The 16 GB configuration has a very narrow VRAM margin. Different drivers, CUDA versions, batch shapes, or background GPU workloads can cause OOM.
Do not run ComfyUI or another heavy CUDA workload at the same time.
Keep the F16 Vision projector on CPU for this profile.
ZeroRefusal/uncensored models can produce unsafe, illegal, or incorrect content. Deployers are responsible for access control, auditing, and safety policy.
Follow LICENSE and all upstream model, dataset, and tool licenses and terms.

QQZ2026/Qwen3.8-27B-ZeroRefusal-UD-IQ4_XS-MTP-GGUF