---

license: apache-2.0

language:

• en
• zh

library_name: gguf

pipeline_tag: image-text-to-text

base_model:

• asfgsdfg/Qwen3.8-27B-Heretic

base_model_relation: quantized

model_name: Qwen3.8-27B-LowDrift-UD-IQ4_XS-MTP-GGUF

tags:

• gguf
• llama-cpp
• qwen
• qwen3_5
• qwen3.8
• vision
• multimodal
• heretic
• lowdrift
• reduced-refusal
• abliterated
• uncensored
• quantized
• iq4_xs
• ud-iq4-xs
• mtp
• speculative-decoding
• unsloth
• dynamic-v3-preview-layout
• rtx-5060-ti
• 16gb-vram

---

Qwen3.8-27B LowDrift UD-IQ4_XS + Embedded MTP GGUF

Low-drift / reduced-refusal Qwen3.8-27B derivative using an earlier Unsloth UD-IQ4_XS preview-era tensor layout and an embedded original MTP layer, packaged as GGUF for llama.cpp.

> Quantization provenance note: the 13.274 GiB UD-IQ4_XS reference artifact used for this build was an earlier Unsloth preview-era layout that is no longer part of the current official Dynamic V3 lineup. This repository does not claim to be the current official Unsloth Dynamic V3 release.

Download

Repository:

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

Download the GGUF with the Hugging Face CLI:

hf download QQZ2026/Qwen3.8-27B-LowDrift-UD-IQ4_XS-MTP-GGUF \
  Qwen3.8-27B-LowDrift-UD-IQ4_XS-MTP.gguf \
  --local-dir .

Or let llama.cpp download the exact file:

llama-server \
  -hf QQZ2026/Qwen3.8-27B-LowDrift-UD-IQ4_XS-MTP-GGUF \
  -hff Qwen3.8-27B-LowDrift-UD-IQ4_XS-MTP.gguf \
  --port 8001

The GGUF is the model artifact. Vision additionally requires a compatible Qwen3.8-27B F16 mmproj as documented below.

Files

Main model:

Qwen3.8-27B-LowDrift-UD-IQ4_XS-MTP.gguf

Integrity:

size    14,252,845,184 bytes (~13.274 GiB)
sha256  49021e6e76af0ac6298e56aa4fab1ed56b62c7c66b6e7a18933907185bd1827d

GGUF verification:

architecture       qwen35
block_count        65
nextn_predict      1
tensor_count       866
main quant class   IQ4_XS - 4.25 bpw
FINAL VERIFY       PASS

The 851 main-model tensor types reproduce the reference Unsloth UD-IQ4_XS layout with 0 mismatch. The 15 blk.64 MTP tensors were grafted byte-for-byte from the reference UD GGUF and verified by SHA256.

Recommended RTX 5060 Ti 16 GB profile

Production target validated on a single RTX 5060 Ti 16 GB:

llama-server \
  -m Qwen3.8-27B-LowDrift-UD-IQ4_XS-MTP.gguf \
  --alias qwen3.8-27b \
  --mmproj mmproj-Qwen3.8-27B-F16.gguf \
  --no-mmproj-offload \
  --image-max-tokens 4096 \
  -c 68000 \
  -np 2 \
  --kv-unified \
  -ngl 999 \
  --flash-attn on \
  -ctk q4_0 \
  -ctv q4_0 \
  --spec-type draft-mtp \
  --spec-draft-n-max 2 \
  --spec-draft-type-k f16 \
  --spec-draft-type-v f16 \
  -b 512 \
  -ub 64 \
  --threads 7 \
  --fit off \
  --jinja

Measured cold-load state at the 68K production point:

runtime n_ctx      68,096
parallel slots     2
GPU used           ~15,650 MiB
GPU free           ~240 MiB

Recommended workload policy:

• short customer-service requests: cap at about 4K context;
• main reasoning/coding requests: cap at about 64K;
• leave the remaining physical context for system prompts, tools and template overhead.

MTP performance observations

Same RTX 5060 Ti 16 GB, same model family and server build:

72K / P2 / MTP-2
  decode               51.48 tok/s
  draft acceptance     75.66%
  mean accepted length 2.51

76K / P2 / MTP-2
  decode               50.78 tok/s
  draft acceptance     73.53%
  mean accepted length 2.47

88K / P2 / MTP-1
  decode               43.28 tok/s
  draft acceptance     86.99%
  mean accepted length 1.87

80K / P2 / MTP-2
  OOM during MTP-context allocation

68K was selected as the production profile to preserve MTP-2 throughput while providing more VRAM margin than 72K.

CUDA pool warm-up

A repeated 24K-prefill diagnostic at 60K / P2 / MTP-2 observed a one-time ~106 MiB retained CUDA-pool allocation on the first long prefill, followed by +0 MiB on the second and third repeats. In that test this behaved like a pool high-water allocation, not a continuing stepwise leak.

Vision

A matching Qwen3.8 F16 mmproj can be used with CPU residency on a 16 GB GPU:

--mmproj mmproj-Qwen3.8-27B-F16.gguf
--no-mmproj-offload
--image-max-tokens 4096

--image-max-tokens 4096 caps the image-token contribution; it does not force total request prefill to exactly 4096 tokens. Text, chat-template, tool-schema and history tokens are additional.

The projector is not duplicated in this repository. It can be reused from a compatible Qwen3.8-27B Vision deployment.

Provenance and license

• Base family: Qwen3.8-27B
• Behavioral derivative: asfgsdfg/Qwen3.8-27B-Heretic
• Quantization-layout reference: earlier Unsloth Qwen3.8-27B UD-IQ4_XS preview-era artifact
• Runtime: llama.cpp, b10435-based FA-transient build used during validation
• License: Apache-2.0, inherited from the behavioral source/base model

This derivative is not mathematically identical to the original model. The behavioral source was selected to reduce refusal behavior while keeping published output-distribution drift low.

Reproducibility

Build notes, production command, checksum and validation data:

https://github.com/wilsonzhang2/qwen3.8-27b-nvfp4-16gb/tree/main/lowdrift-ud-iq4xs-mtp