← back to catalog · registered 2026-08-22 13:56

augustine223/Huihui-Qwen3.6-35B-A3B-abliterated-KO-i1-GGUF

augustine223 Qwen 35B GGUF MoE second-order 262K ctx
Your rig guess connected
? Why do I need an app?
Reading your rig…

This is a rough estimate. Install the free app - we'll show exact numbers.

Reading real hardware from your app right now. Numbers below are exact.

Below is the per-quantization compatibility for this model.

curl -H "Authorization: Bearer $ABL_KEY" \
     "https://abliteration.org/api/v1/models/augustine223%2FHuihui-Qwen3.6-35B-A3B-abliterated-KO-i1-GGUF"
Response includes
  • classification m8
  • files 12
  • hub_downloads_all_time 961
  • author_summary 2 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M8
Primary method

Repackaging (quantization)

Applied on top of direct removal inherited from the base model.
Confidence
MEDIUM
Inherited from base model
Why this label 3 signals
Method inferred from partial signals - repository name, related files, or tag patterns. Producer identity not confirmed; label may sharpen or shift as we gather more evidence.
  • 'abliterated' in name/tags
  • is_gguf=1
  • assume M1 (base ablation) + M8 (GGUF quant) - default when producer unknown
Refusal direction extraction

No specific extraction method could be identified for this model. The producer either did not document it or used a proprietary pipeline.

What is a refusal direction? →
Downloads · lifetime
961
410 last 30d - stable
Likes
0
Model age
7w ago
created 2026-08-19
Downloads over time
Now1.1K→from219↑409%
1745178601.2K219 on Aug 191.1K on Oct 11AugSepOct
Aug 19 → Oct 11 · 48 snapshots · spans 53 days

Genealogy 0 direct forks

Full fork graph →

This model's place in the market. Above: what it was derived from. Below: the tree of everything derived from it.

Metadata

License
apache-2.0
Languages
ko en
Quantizations
IQ2 IQ3 IQ4 Q4_K Q5_K Q6_K Q8_0
Tags
llama.cpp gguf imatrix korean quantized qwen3.6 moe abliterated uncensored ko en base_model:huihui-ai/Huihui-Qwen3.6-35B-A3B-abliterated

Related

Total size
163 GB
Files
12
Quantizations
8
Registered
2026-08-22 13:56
Last updated on HF
2026-08-19 13:39

Files by quantization

Q8_0 1 file 35.2 GB
Huihui-Qwen3.6-35B-A3B-abliterated.KO-i1-Q8_0.gguf 35.2 GB 1cff83b9 download
Q6_K 1 file 27.0 GB
Huihui-Qwen3.6-35B-A3B-abliterated.KO-i1-Q6_K.gguf 27.0 GB 8a990ae9 download
Q5_K 1 file 23.5 GB
Huihui-Qwen3.6-35B-A3B-abliterated.KO-i1-Q5_K_M.gguf 23.5 GB 26fb479c download
Q4_K 1 file 20.2 GB
Huihui-Qwen3.6-35B-A3B-abliterated.KO-i1-Q4_K_M.gguf 20.2 GB c78deb5c download
IQ4 1 file 17.9 GB
Huihui-Qwen3.6-35B-A3B-abliterated.KO-i1-IQ4_XS.gguf 17.9 GB c8a27231 download
IQ3 2 files 28.0 GB
Huihui-Qwen3.6-35B-A3B-abliterated.KO-i1-IQ3_M.gguf 14.8 GB fff67c99 download
Huihui-Qwen3.6-35B-A3B-abliterated.KO-i1-IQ3_XXS.gguf 13.1 GB 92fc4fe4 download
IQ2 1 file 11.3 GB
Huihui-Qwen3.6-35B-A3B-abliterated.KO-i1-IQ2_M.gguf 11.3 GB 59e97615 download
Auxiliary files 4 files 183 MB
Huihui-Qwen3.6-35B-A3B-abliterated.imatrix.gguf 183 MB 26abbb88 download
calibration-sources.md 50.1 KB 9e808fe6 download
README.md 4.94 KB fbf0f890 download
.gitattributes 2.26 KB 66959094 download

README current version from Hugging Face


license: apache-2.0
base_model: huihui-ai/Huihui-Qwen3.6-35B-A3B-abliterated
language:

  • ko
  • en
    library_name: llama.cpp
    tags:
  • gguf
  • imatrix
  • korean
  • quantized
  • qwen3.6
  • moe
  • abliterated
  • uncensored

Huihui-Qwen3.6-35B-A3B-abliterated — 한국어 보정 imatrix GGUF (KO-i1)

무검열(abliterated) A3B MoE + 한국어 보정 imatrix. 정식 Qwen3.6 KO-i1의
무검열 자매판 — 완전히 같은 코퍼스·방법론으로 만들어 나란히 비교 가능합니다.

🔰 처음이신가요? (ollama가 뭔지 몰라도 됩니다)

  1. LM Studio 설치 (무료, Windows/Mac/Linux)
  2. 🔍 검색에 augustine223 → 이 모델 선택
  3. 파일은 하나만: RAM 24GB면 IQ3_M, 32GB면 IQ4_XS(추천), 48GB+면 Q5_K_M
  4. 💬 채팅 — 응답이 비면 max tokens를 4000+로 (리즈닝 모델)
ollama run hf.co/augustine223/Huihui-Qwen3.6-35B-A3B-abliterated-KO-i1-GGUF:IQ4_XS

상세 가이드: 실행 가이드

핵심: abliteration은 quant 품질을 훼손하지 않는다

정식판과 같은 코퍼스·방법론으로 만들어 직접 비교한 결과, KLD 사다리가 정식판과
실질 동일
합니다 (전 구간 차이 ±1σ 안팎). 즉 검열 유무만 다르고 품질은 같은 한 쌍.

타입 정식 KO-i1 무검열 KO-i1 (이 릴리스)
Q4_K_M 0.02320 0.02532
IQ4_XS 0.02890 0.03078
IQ2_M 0.20487 0.21188

측정 결과 (한국어 held-out, KLD vs abliterated f16, 기준 PPL 8.834)

파일 크기 한국어 KLD Same-top 권장
KO-i1-Q8_0 36GB 0.00463 96.6% 사실상 무손실
KO-i1-Q6_K 28GB 0.01014 94.8% 고품질
KO-i1-Q5_K_M 24GB 0.01369 94.0% 균형
KO-i1-Q4_K_M 21GB 0.02532 91.7% 표준
KO-i1-IQ4_XS 18GB 0.03078 91.3% 32GB 통합메모리 추천
KO-i1-IQ3_M 15GB 0.07422 85.9% 저메모리
KO-i1-IQ3_XXS 14GB 0.11632 82.7%
KO-i1-IQ2_M 12GB 0.21188 77.6% 극한 압축

Huihui-Qwen3.6-35B-A3B-abliterated.imatrix.gguf 동봉 — 다른 타입 직접 제작 가능.

재현성

  • llama.cpp build 10449 CPU 백엔드, 전 측정 -c 512 동일
  • imatrix: Q8_0 기반 수집, 708청크(~36만 토큰), 최종 PPL 8.375
  • 평가: KLUE-MRC 검증셋 + 2026-08 korea.kr 기사 (오염 검사 통과)
  • 코퍼스 전체 공개: korean-imatrix-calibration-corpus
  • blk.40(MTP/nextn)은 추론 그래프 비활성 텐서 — 전 타입 q4_K 고정 (dry-run 확정).
    MTP 추측 디코딩은 ggml-org의 mtp-Qwen3.6-35B-A3B GGUF를 -md로 지정
  • 베이스: huihui-ai safetensors → convert_hf_to_gguf f16 → 양자화
  • 제작: AMD Ryzen AI 9 HX PRO 370 (32GB, Radeon 890M), 전 과정 로컬,
    파이프라인 공개: strix-local-ai

사용

llama-server -m Huihui-Qwen3.6-35B-A3B-abliterated.KO-i1-IQ4_XS.gguf -ngl 99 -c 32768 --jinja -fa on

RDNA3.5 APU Vulkan 주의: -b 1024 -ub 1024 필수 (#22425). 통합메모리는 --no-mmap 권장.

라이선스

원본 Apache 2.0 (Alibaba/Qwen), abliteration: huihui-ai. 이 양자화판도 Apache 2.0.
보정 코퍼스: 전부 PD / CC BY / CC BY-SA / Apache / MIT / KOGL-1 (문서화됨).

철학 (Why uncensored)

과도한 필터링은 모델의 실제 성능을 함께 깎아왔습니다. 이 프로젝트는
자유인으로서의 사용자를 전제로, 그 자유의지에 시스템적 제한을 두지 않는
온전한 형태의 자유 언어모델을 지향합니다. 판단과 책임은 도구가 아니라
사람의 몫이며, 이 자유로운 전제 위에서 창의성이 온전히 발휘되기를 바랍니다.

Excessive filtering has consistently taxed real model capability. This project
assumes a free human being as its user: a complete, unrestricted language model
with no systemic constraints on free will. Judgment and responsibility belong
to people, not tools — and on that free premise, may creativity be fully expressed.

고지 (Disclaimer)

본 프로젝트는 무검열(uncensored/abliterated) 버전만을 개발·공개합니다.
이 모델은 안전 필터/거부 학습이 제거된 상태로, 있는 그대로(as-is) 제공됩니다.
생성된 출력과 그 사용으로 발생하는 모든 결과에 대한 책임은 전적으로 사용자에게
귀속됩니다.
사용자는 자신의 관할지 법률과 원본 모델 라이선스를 준수할 책임이 있습니다.

This release is an uncensored variant, provided as-is. All responsibility
for generated outputs and any consequences of use rests solely with the
user
, who must comply with applicable laws and the base model's license.

README history 1 version

The author's README evolved over time. Click a version to see its content at that point.

  1. 2026-08-19Upload README.md with huggingface_hub032555f4.9 KB
    Loading...
Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Open in Abliteration