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.
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.
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.
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.
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.
