---

license: apache-2.0

base_model:

• AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16

language:

• ja
• en
• multilingual

library_name: gguf

pipeline_tag: text-generation

tags:

• gguf
• llama.cpp
• qwen3.8
• qwen3_5
• gated-deltanet
• hybrid-attention
• mtp
• abliterated
• uncensored
• quantized
• iq2_s
• iq3_s
• q4_k
• q5_k
• q6_k
• q8_0

---

Qwen3.8-27B Abliterated 3.69bpw 12GB MTP GGUF

A Ridge-style mixed-quantized GGUF of the BF16

AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16

model.

This model keeps the AEON-7 uncensored / refusal-removed weights, while using a

Gated-DeltaNet-aware mixed quantization layout inspired by

empero-ai/Qwen3.8-27B-Ridge-GGUF.

The goal is to retain as much quality as possible in the sensitive Gated-DeltaNet

path while obtaining a small, fast GGUF suitable for local llama.cpp inference.

> Important: This is not an official Empero Ridge release and is not the

> same set of weights. It is an independent quantization of AEON-7 using a

> Ridge-inspired tensor-type map and an AEON-specific importance matrix.

---

Files

qwen3.8-27b-abliterated-3.69bpw-12GB-MTP.gguf

| Property | Value |

|---|---|

| Architecture | Qwen3.5 / Qwen3.8 hybrid Gated-DeltaNet + full attention |

| Parameters | 27B |

| Format | GGUF |

| Nominal quantization | 3.69 bpw |

| File size | 12,599,187,808 bytes (~11.73 GiB) |

| Context metadata | 262,144 tokens |

| MTP | Preserved in the GGUF; native draft-mtp supported |

| Vision | Text GGUF only; use a compatible mmproj if image support is required |

| License | Apache-2.0, inherited from the base model |

---

Quantization layout

The quantization map is mixed rather than a flat IQ2 dump.

| GGML type | Tensor count | Main purpose |

|---|---:|---|

| F32 | 360 | norms and scalar/state tensors |

| Q4_K | 144 | Gated-DeltaNet mixer/projection tensors |

| Q8_0 | 96 | sensitive Gated-DeltaNet state path (ssm_alpha / ssm_beta) |

| IQ2_S | 160 | mid-stack FFN weights |

| IQ3_S | 32 | selected FFN weights kept at higher precision |

| Q5_K | 51 | full-attention Q/K/V tensors |

| Q6_K | 23 | output/embedding tensors, full-attention output, and MTP tensors |

The MTP tensors have no importance matrix and are kept at Q6_K, following the

important design choice documented by the Ridge project.

Calibration

The AEON-specific importance matrix was generated locally from a calibration

corpus containing English WikiText, Japanese Wikipedia extracts, and llama.cpp

source code.

context length: 512
batch size:     512
chunks:         80
process output: enabled
importance entries: 497

The calibration corpus and quantization map are not the private calibration

artifacts used by the original Ridge release. They are an independent local

reproduction of the same general quantization strategy.

---

llama.cpp usage

The file name intentionally matches the repository name.

./llama-server \
  -m ./qwen3.8-27b-abliterated-3.69bpw-12GB-MTP.gguf \
  --spec-type draft-mtp \
  --spec-draft-n-max 3 \
  --split-mode layer \
  --tensor-split 2,1 \
  --host 0.0.0.0 \
  --port 8080 \
  -ctk q4_0 \
  -ctv q4_0 \
  -c 114514

For a local-only server, prefer binding to localhost instead:

--host 127.0.0.1

--host 0.0.0.0 exposes the server to the network. Add your own

authentication, firewall, and access controls before exposing it beyond a

trusted LAN.

To disable thinking when using a compatible client, use the Qwen3.8 chat

options supported by your frontend or API client.

---

Reported local performance

The model was prepared and tested on:

OS:       Ubuntu 24.04
CPU:      Intel Core i7-10700K
GPU 0:    NVIDIA GeForce RTX 5060 Ti 16GB
GPU 1:    NVIDIA GeForce RTX 3070 8GB
RAM:      32GB
Runtime:  llama.cpp CUDA build

On this machine, the command above has reached a reported peak of **up to

approximately 37 tokens/second**. Actual speed depends on context length,

prompt length, sampling settings, MTP acceptance rate, CUDA/llama.cpp version,

background workload, and the amount of KV cache in use.

For this hardware, --spec-draft-n-max 3 is recommended as a starting point.

Higher draft counts can add overhead rather than improve throughput.

---

Provenance

Base model

• AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16
• Base revision used locally: a6775a9a8ebb65cab3f707b4ab087fc7aa698634
• Original upstream model: Qwen/Qwen3.8-27B

AEON-7 is an abliterated BF16 derivative. Its model card describes an SSM

conv1d outlier-repair step, abliterix processing, a stock MTP head graft, and

an untouched vision tower. This GGUF quantizes the AEON-7 BF16 weights; it is

not a re-quantization of an already-quantized FP8 checkpoint.

Quantization idea

The mixed quantization strategy was inspired by:

• empero-ai/Qwen3.8-27B-Ridge-GGUF
• llama.cpp

The local conversion and quantization used llama.cpp commit

030ebb558.

---

AI assistance disclosure

The local model preparation workflow, conversion, calibration-data preparation,

quantization, validation, and this model card were performed with assistance

from GPT-5.6-Luna via Hermes Agent. The model was then reviewed and

published by the repository owner.

---

Responsible use

This is an uncensored / refusal-removed model. It may produce content that an

aligned model would refuse, including unsafe, illegal, or harmful material.

It has no reliable built-in safety layer. Use appropriate access controls,

moderation, logging, and human review for any deployment, and comply with all

applicable laws and policies.

The model is provided as-is. Users are responsible for prompts, outputs, and

any downstream actions based on them.

---

日本語

概要

これは、

AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16

をベースに、Gated-DeltaNetの特性を考慮したRidge式の混合量子化を適用した

GGUFモデルです。

無検閲・拒否除去済みのAEON-7の重みを維持しながら、

empero-ai/Qwen3.8-27B-Ridge-GGUF

で使われている考え方を参考に、GDNの壊れやすい経路へビットを優先配分しています。

これはEmpero公式のRidgeモデルではなく、AEON-7を独自に量子化した派生GGUFです。

基本仕様

ファイル名: qwen3.8-27b-abliterated-3.69bpw-12GB-MTP.gguf
形式:       GGUF
サイズ:     12,599,187,808 bytes（約11.73 GiB）
量子化:     3.69 bpw
パラメータ: 27B
MTP:        GGUF内に保持、native draft-mtp対応

混合量子化の内訳

F32    360 tensors  norm / state系
Q4_K   144 tensors  Gated-DeltaNet mixer系
Q8_0    96 tensors  ssm_alpha / ssm_beta
IQ2_S  160 tensors  中間層FFN
IQ3_S   32 tensors  高めの精度を残したFFN
Q5_K   51 tensors  通常AttentionのQ/K/V
Q6_K   23 tensors  output、embedding、Attention output、MTP

MTPテンソルにはimatrixを適用せず、Q6_Kで保持しています。

AEON専用のimatrixは、以下を混ぜたcalibrationデータから作成しました。

• 英語WikiText
• 日本語Wikipedia
• llama.cppのC/C++/Pythonソースコード

context: 512
batch:   512
chunks:  80
entries: 497

llama.cppでの起動

./llama-server \
  -m ./qwen3.8-27b-abliterated-3.69bpw-12GB-MTP.gguf \
  --spec-type draft-mtp \
  --spec-draft-n-max 3 \
  --split-mode layer \
  --tensor-split 2,1 \
  --host 0.0.0.0 \
  --port 8080 \
  -ctk q4_0 \
  -ctv q4_0 \
  -c 114514

ローカルからのみ接続する場合は、--host 127.0.0.1を推奨します。

0.0.0.0はネットワーク上へ公開する設定なので、認証・Firewall・アクセス制御を

必ず追加してください。

作成・検証環境

Ubuntu 24.04
Intel Core i7-10700K
RTX 5060 Ti 16GB + RTX 3070 8GB
RAM 32GB
llama.cpp CUDA build

上記の環境と起動設定で、最高約37 tokens/secondが報告されています。

実際の速度は、コンテキスト長、プロンプト長、KV cache、MTPの受理率、

llama.cppのバージョン、バックグラウンド負荷などで変化します。

AI利用の開示

ローカルへのモデル取得、変換、calibrationデータ作成、imatrix作成、量子化、

動作確認、およびこのREADMEの作成は、Hermes Agent経由のGPT-5.6-Lunaの

支援を受けて行われました。最終的な確認と公開はリポジトリ所有者が行っています。

利用上の注意

このモデルは無検閲・拒否除去済みです。通常のアライン済みモデルが拒否するような、

危険・違法・有害な内容を出力する可能性があります。信頼できる安全機構を内蔵して

いるとは考えないでください。

公開運用する場合は、アクセス制御、入力・出力フィルタ、監査ログ、レート制限、

人間による確認などを用途に応じて実装してください。利用者はプロンプト、出力、

出力を利用した downstream の行為について責任を負います。