license: apache-2.0
language:

• ko
• en
  library_name: llama.cpp
  pipeline_tag: text-generation
  inference: false
  base_model:
• huihui-ai/Huihui-Qwen3.8-27B-abliterated
  tags:
• gguf
• llama.cpp
• qwen3.8
• korean
• abliterated
• quantized
• imatrix
• ridge
• dflash2
• speculative-decoding

Huihui-Qwen3.8-27B-abliterated KO Ridge 3.7bpw GGUF

A 3.69 BPW mixed-tensor GGUF tuned to run Huihui Qwen3.8 27B Abliterated with a Qwen3.8 DFlash 2 draft on an NVIDIA RTX 5060 Ti 16 GB using llama.cpp.

이 저장소는 RTX 5060 Ti 16GB 한 장에서 Huihui Qwen3.8 27B Abliterated와 DFlash 2를 함께 실행하기 위한 3.69 BPW 혼합 텐서 GGUF입니다. 한국어/영어 importance matrix와 Ridge 계열 텐서 배치를 사용했습니다.

Files

The DFlash 2 draft is not duplicated here. Download Qwen3.8-27B-DFlash2-Q4_K_M.gguf from z-lab/Qwen3.8-27B-DFlash2-GGUF.

Important quantization note

This model was requantized from the KO-i1 Q8_0 GGUF, not directly from BF16. Requantization can lose more quality than a single BF16-to-GGUF conversion. The source Q8_0 and Korean/English importance matrix come from augustine223/Huihui-Qwen3.8-27B-abliterated-KO-i1-GGUF.

The tensor layout follows the size/quality strategy of empero-ai/Qwen3.8-27B-Ridge-GGUF:

The resulting target is 12,599,186,976 bytes, 12,005.04 MiB, and 3.69 BPW.

SHA-256:

4f72d4e3b6723fb259e7f8244c70fbf2850151aab3745d8c1e84768fba161515

Tested RTX 5060 Ti 16 GB profile

Tested locally with a Blackwell RTX 5060 Ti, CUDA 13.1, a DFlash 2-enabled llama.cpp build, and the Q4_K_M DFlash draft.

Short smoke-test results were approximately 44.7 tok/s on the first generation and 62.6 tok/s on a cached short request. These are not full benchmark results; prompt length, sampler, CPU, driver, and draft acceptance change throughput.

If your system has less free VRAM, lower CTX first, for example CTX=65536 or CTX=32768.

1. Download

Install the Hugging Face CLI if needed, then download the target and draft into one directory:

python3 -m pip install -U "huggingface_hub[hf_xet]"

hf download \
  donghanasd/Huihui-Qwen3.8-27B-abliterated-KO-Ridge-3.7bpw-GGUF \
  Huihui-Qwen3.8-27B-abliterated-KO-Ridge-3.7bpw.gguf \
  --local-dir .

hf download \
  z-lab/Qwen3.8-27B-DFlash2-GGUF \
  Qwen3.8-27B-DFlash2-Q4_K_M.gguf \
  --local-dir .

2. Build llama.cpp with DFlash 2

DFlash 2 support is available through llama.cpp PR #27342. The following is the tested NVIDIA build shape:

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
git fetch origin pull/27342/head:pr-27342
git switch pr-27342

cmake -S . -B build -G Ninja \
  -DCMAKE_BUILD_TYPE=Release \
  -DGGML_CUDA=ON \
  -DGGML_CUDA_FA=ON \
  -DGGML_CUDA_GRAPHS=ON \
  -DCMAKE_CUDA_ARCHITECTURES=120

cmake --build build --target llama-server -j

If the PR has already been merged into your llama.cpp version, a current checkout with --spec-type draft-dflash support can be used instead.

3. Run llama-server

Download run-rtx5060ti.sh, place it next to both GGUF files, and point it at the built server:

chmod +x run-rtx5060ti.sh
LLAMA_SERVER=/path/to/llama.cpp/build/bin/llama-server ./run-rtx5060ti.sh

Equivalent core command:

GGML_CUDA_PDL=0 \
GGML_CUDA_ENABLE_UNIFIED_MEMORY=1 \
/path/to/llama-server \
  --host 127.0.0.1 --port 8080 \
  --model Huihui-Qwen3.8-27B-abliterated-KO-Ridge-3.7bpw.gguf \
  --model-draft Qwen3.8-27B-DFlash2-Q4_K_M.gguf \
  --jinja --no-mmproj \
  -ngl 99 --spec-draft-ngl 99 \
  --spec-type draft-dflash \
  --spec-draft-n-max 4 --spec-draft-n-min 1 \
  -fa on --ctx-size 131072 \
  --cache-type-k q4_0 --cache-type-v q4_0 \
  --spec-draft-type-k q4_0 --spec-draft-type-v q4_0 \
  --parallel 1 --kv-unified --fit off --no-context-shift \
  --batch-size 512 --ubatch-size 128 \
  --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0 \
  --reasoning on --reasoning-budget -1 --reasoning-preserve

Health check:

curl http://127.0.0.1:8080/health

OpenAI-compatible request:

curl http://127.0.0.1:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "huihui-qwen3.8-27b-ko-ridge",
    "messages": [{"role": "user", "content": "안녕하세요!"}],
    "max_tokens": 512,
    "temperature": 1.0,
    "top_p": 0.95
  }'

This is a reasoning model. Small max_tokens values can be consumed entirely by reasoning_content, leaving content empty. Use a larger limit for normal chat.

Security

The supplied launcher binds to 127.0.0.1 by default. It refuses a non-loopback bind unless API_KEY is set. Do not expose an unauthenticated llama-server directly to the internet.

Example LAN bind with authentication:

HOST=0.0.0.0 API_KEY='replace-with-a-strong-secret' \
LLAMA_SERVER=/path/to/llama-server ./run-rtx5060ti.sh

Do not commit API keys or tokens to a repository.

Compatibility and limitations

• This release is text-only in the tested configuration (--no-mmproj).
• The final MTP/next-token-prediction tensors are retained in the GGUF, but the tested target graph ignores them while the external DFlash 2 draft handles speculative decoding.
• DFlash 2 verifies draft tokens against the target model. Acceptance and speed can differ from the base Qwen checkpoint because this target is an abliterated derivative.
• The model is an abliterated/uncensored derivative. It may generate unsafe, inaccurate, biased, or objectionable content. Users are responsible for lawful and appropriate use.
• The 262K trained-context metadata does not mean 262K will fit on a 16 GB GPU. The validated 16 GB profile is 128K with quantized KV caches and one slot.

Provenance and credits

• Base architecture: Qwen/Qwen3.8-27B
• Abliterated model: huihui-ai/Huihui-Qwen3.8-27B-abliterated
• KO-i1 Q8_0 and importance matrix: augustine223/Huihui-Qwen3.8-27B-abliterated-KO-i1-GGUF
• Ridge tensor-layout reference: empero-ai/Qwen3.8-27B-Ridge-GGUF
• DFlash 2 GGUF: z-lab/Qwen3.8-27B-DFlash2-GGUF
• DFlash project: z-lab/dflash
• Runtime: ggml-org/llama.cpp

All referenced model repositories report Apache-2.0 licensing. This repository is released under Apache-2.0. Follow the licenses and terms of every upstream component.

DFlash 2 citation

@misc{inco2026dflash2,
  title  = {{DFlash 2: Keep Drafting Parallel}},
  author = {{Inco AI}},
  year   = {2026},
  month  = {August},
  url    = {https://inco.ai/blog/dflash2/}
}