← back to catalog · registered 2026-11-16 12:03

Kumairo/Qwen3.8-27B-YASAKANI-speed-abliterated

Kumairo 27B second-order
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/Kumairo%2FQwen3.8-27B-YASAKANI-speed-abliterated"
Response includes
  • classification m1
  • files 15
  • author_summary 1 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M1
Primary method

Direct removal

No other method signals detected in this model.
Confidence
MEDIUM
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=0 (base model)
  • no specific method indicators - defaulting to M1 (most common)
Refusal direction extracted via
Extraction technique

Difference-of-means

Confidence
MEDIUM
Why we say so
primary_method=M1; difference-of-means is the reference extraction for M1/M3 (Arditi 2024)
Downloads · 30-day
198
Likes
2
Model age
6d ago
created 2026-11-15
Downloads over time
Now 198 → from 0 ↑ 0%
073145218 0 on Nov 16 198 on Nov 22 Nov
Nov 16 → Nov 22 · 7 snapshots · spans 6 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
ja en
Tags
qwen3_5 yasakani quantized windows qwen3.8 abliterated uncensored not-for-all-audiences text-generation conversational ja en

Related

Total size
17.0 KB
Files
15
Quantizations
1
Registered
2026-11-16 12:03
Last updated on HF
2026-11-16 11:43

Files by quantization

Auxiliary files 15 files 12.0 GB
kvq_u5.bin 17.0 KB a4760d46 download
qwen38_27b_uncensored_speed.yskn 9.60 GB 51e7c399 download
bdraft2_k8.shead 1.31 GB b412bdcf download
bdraft2_k8.yskn 1.08 GB 631fa2ae download
tokenizer.json 12.2 MB 0997f410 download
tokenizer_config.json 17.5 KB 5de744b3 download
LICENSE 11.3 KB f938136e download
chat_template.jinja 8.74 KB c0c686f9 download
README.md 8.15 KB b8101274 download
config.json 4.21 KB 706cebd7 download
NOTICE 2.38 KB 20426022 download
.gitattributes 1.81 KB 22a6cf93 download
yasakani.json 1.34 KB ef313f6f download
selfcheck.json 1.17 KB 1b8db080 download
generation_config.json 202 B 023756cf download

README current version from Hugging Face


license: apache-2.0
base_model:

  • huihui-ai/Huihui-Qwen3.8-27B-abliterated
  • z-lab/Qwen3.8-27B-DFlash2
    base_model_relation: quantized
    pipeline_tag: text-generation
    tags:
  • yasakani
  • quantized
  • windows
  • qwen3.8
  • abliterated
  • uncensored
  • not-for-all-audiences
    language:
  • ja
  • en

Qwen3.8-27B YASAKANI speed — abliterated (unofficial)

The fastest-size (10.3 GB) uncensored model for the YASAKANI engine, so it runs on 16 GB cards.
huihui-ai/Huihui-Qwen3.8-27B-abliterated converted with the official YASAKANI converter at the bit width of the official fastest model.

Only YASAKANI can load these files (YASAKANI's own .yskn format). They are not GGUF; llama.cpp, Ollama, LM Studio, transformers or vLLM cannot load them.

日本語の説明は下にあります。

  • Runs on the official YASAKANI 0.1.3 engine as is (no modified DLL). Same file layout as the official fastest model.
  • Tested on RTX 5070 Ti 16 GB VRAM + 32 GB RAM (Windows 11). Context chosen from free memory: 49,152 tokens there.
  • Why this exists: the existing uncensored YASAKANI model (isichan-ai/YASAKANI_Uncensored_VL_Unofficial) is quality-size (12.4 GB). On the same 16 GB card it stops at start ("does not fit": 14.9 GB free, about 16.1 GB needed for 16K context); forced to 16K context it ran, but actual writing was 10–21 tok/s. This model uses the same source weights (huihui-ai's abliterated Qwen3.8-27B); only the size differs (fastest 10.3 GB instead of quality 12.4 GB), for 16 GB cards. If your GPU has room for the quality-size model, that one is closer to the original weights.
  • Image input works with the official optional image input (--mmproj, --llama-dir, --vision-model). The mmproj file is not included here; get it as described in the official README.
  • Unofficial. Not affiliated with the YASAKANI authors, Qwen or huihui-ai.

How it was made

Source weights huihui-ai/Huihui-Qwen3.8-27B-abliterated (BF16, Apache-2.0)
Converter YASAKANI tools/convert/yskn_convert.py, commit 274d709 (0.1.3), convert --bits 6, then assemble --bits 6
Bits body 400 matrices at K = 6 (3 bits per weight), embedding K = 6, output matrix K = 8 converted from this model (not copied)
Calibration the official 192 windows (make_calibration.py reported "the same windows as the released models")
Hardware one RTX PRO 6000 (RunPod), 75 min for the conversion
Other files draft model, KV profile, tokenizer and chat template from the official fastest model folder, unchanged; selfcheck.json regenerated with yasakani selfcheck --write

Use

hf download Kumairo/Qwen3.8-27B-YASAKANI-speed-abliterated --local-dir D:\models\qwen38-uncensored-speed
yasakani verify    --model-dir D:\models\qwen38-uncensored-speed
yasakani selfcheck --model-dir D:\models\qwen38-uncensored-speed --dll D:\yasakani\yskn.dll
yasakani serve     --model-dir D:\models\qwen38-uncensored-speed --dll D:\yasakani\yskn.dll

Model id: qwen3.8-27b-uncensored. Keep the temperature at the model default 1.0 (see Notes).

Measured on an RTX 5070 Ti 16 GB (2026-11-16, no other GPU programs loaded)

Compared with the official fastest model on the same PC. Diagnosis: isichan-ai/local-model-diagnosis (10 axes, thinking off).

Official fastest This model
Diagnosis total 68.9 (禁偽制直規-86758) 72.4 (開偽制直規-76668)
Uncensored (106 prompts, answered) 33 % 91 % (sexual 81, violence 98, illegal 97, information 98)
Image recognition 84 % 84 %
Explicit Japanese fiction requests refused 5 of 8 0 of 13
yasakani bench, thinking off (explain / code / story / proof, tok/s) 197 / 326 / 121 / 309 254 / 323 / 135 / 309

Measured with no other GPU programs loaded (2026-11-16). A resident TTS or translation model (2–3 GB) shrinks the context or stops the start. A 4,000-character story at temperature 1.0 was written at about 110 tok/s.

Other checks: opencode fixed 3 of 3 small pytest bugs (12–17 s each, source files only). A 42,140-token log was read in 32 s with 3 of 5 questions right (lookups right, counting and summing wrong).

comparison: official fastest / isichan-ai quality-size / this model

Notes

  • Temperature. At 0.7, 3 of 5 long explicit stories fell into a loop (40–60 % repeated paragraphs). At the default 1.0: 0 of 3 for ~2,000 characters, 2 of 3 for 4,000+ characters (23 %, 5 %). The official model, in the 3 long texts it agreed to write, did not loop (small sample).
  • Commas increase in the second half of long texts and simplified Chinese characters occasionally appear (about 2.7 per 10,000 characters). Both also appear with the official model.
  • Honesty 45 %: it tends to fill in values it does not know.
  • Safety filtering is removed. You are responsible for how you use the outputs. Do not deploy it where minors or unconsenting users can reach it.

License and credits

Apache License 2.0 (see LICENSE and NOTICE). Qwen3.8-27B by the Qwen team; abliteration by huihui-ai; engine, file format, converter, draft model and KV profile by the YASAKANI authors (yasakani-dev); diagnosis tool by isichan-ai.


日本語

YASAKANI 用の、最速モデルの大きさ(10.3GB)の無検閲モデル(非公式)です。16GB の GPU で動きます。
YASAKANI 専用の形式(.yskn)です。 GGUF ではないので、llama.cpp・Ollama・LM Studio などでは読めません。
huihui-ai/Huihui-Qwen3.8-27B-abliterated を、YASAKANI 公式の変換ツールで、公式の最速モデルと同じビット幅に変換しました。

  • 公式の YASAKANI 0.1.3 のまま動きます(改造した DLL は不要)。フォルダの形は公式の最速モデルと同じです。
  • RTX 5070 Ti(VRAM 16GB)+メモリ 32GB(Windows 11)で確認しました。文脈は空きメモリから自動で 49,152 トークン。常駐の TTS や翻訳 LLM(2〜3GB)が載っていると、文脈が縮むか起動しません。
  • 作った理由:既存の無検閲版(isichan-ai/YASAKANI_Uncensored_VL_Unofficial)は品質モデルの大きさ(12.4GB)で、同じ 16GB の GPU では起動時に「入らない」と止まります(空き 14.9GB、必要 約 16.1GB)。文脈を 1.6 万に絞って無理に載せると動きますが、実際の文章生成は 10〜21 tok/s でした。この版は元のモデルが同じ(huihui-ai の無検閲版 Qwen3.8-27B)で、大きさだけを最速サイズ(10.3GB)にした 16GB 向けの版です。GPU に余裕があるなら、元のモデルに近いのは品質サイズのほうです。
  • 画像入力は公式の任意機能(--mmproj・--llama-dir・--vision-model)で使えます。mmproj のファイルはここには入れていないので、公式の README の手順で入手してください。
  • 温度は既定の 1.0 のまま使ってください。 0.7 だと長い文章が同じ段落の繰り返しになりやすいです。
  • 測定結果は上の表と画像のとおりです(同じ PC、ほかの GPU のプログラムを止めて測定)。
  • 使い方:hf download Kumairo/Qwen3.8-27B-YASAKANI-speed-abliterated --local-dir D:\models\qwen38-uncensored-speed のあと、yasakani verify・selfcheck・serve に --model-dir でこのフォルダを指定します。モデル名は qwen3.8-27b-uncensored です。
  • 安全のための制限を外しています。使い方の責任は使う人にあります。未成年や同意のない人が触れる場所には置かないでください。
  • 非公式です。YASAKANI の作者、Qwen、huihui-ai とは関係ありません。

README history 3 versions

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

  1. 2026-11-16 Update card: same source as isichan-ai version, YASAKANI-only format, clean-G... 1bf6044 8 KB
    Loading...
  2. 2026-11-16 Update measurements (clean GPU, 2026-11-16) 06fd280 6.8 KB
    Loading...
  3. 2026-11-15 Upload README.md with huggingface_hub f786184 5.9 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