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:
The local conversion and quantization used llama.cpp commit030ebb558.
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。