license: gemma
library_name: gguf
pipeline_tag: text-generation
base_model: AEON-7/Gemma-4-12B-it-AEON-Abliterated-K4-BF16
model_type: gemma4_unified
language:
- en
tags:
Model family
- gemma4
- gemma
- gemma-4-12B
Format
- gguf
- quantization
Abliteration / Uncensored
- uncensored
- abliterated
- unfiltered
- refusal-removed
- biprojection
- multi-direction-biprojection
- k4-biprojection
- aeon
- aeon-7
Capabilities
- text-generation
- chat
- instruct
- reasoning
- coding
- tool-calling
- function-calling
Gemma-4-12B-it AEON Abliterated — K=4 GGUF Quants
This repository contains official GGUF quantizations of AEON-7/Gemma-4-12B-it-AEON-Abliterated-K4-BF16.
The base model is an abliteration of google/gemma-4-12B-it using a custom K=4 multi-direction norm-preserving biprojection that extends standard biprojection recipes with a K-dim orthonormal basis from the top-K SNR layers. This workflow preserves generative quality, slashes wikitext PPL drift compared to K=1 methods, and completely eliminates standard refusal patterns.
Provided Quantization Tiers
| File Name | Quant Type | Size | VRAM / Hardware Recommendation |
|---|---|---|---|
Gemma-4-12B-it-AEON-Abliterated-Q3_K_M.gguf |
3-bit | 6.09 GB | Low-VRAM setups / CPU+GPU split |
Gemma-4-12B-it-AEON-Abliterated-Q4_K_M.gguf |
4-bit | 7.38 GB | Recommended balance. Fits entirely on single T4/RTX 3060/4060 |
Gemma-4-12B-it-AEON-Abliterated-Q5_K_M.gguf |
5-bit | 8.55 GB | Low quality degradation, tight fit on 8GB VRAM |
Gemma-4-12B-it-AEON-Abliterated-Q6_K.gguf |
6-bit | 9.79 GB | High fidelity, excellent for 12GB+ VRAM |
Gemma-4-12B-it-AEON-Abliterated-Q8_0.gguf |
8-bit | 12.70 GB | Near-identical to BF16 precision. Fits comfortably on 16GB VRAM |
Deployment & Usage
1. Python (llama-cpp-python)
To run inference with full GPU acceleration, compile with the CUDA backend and load all layers into VRAM:
llama-cli \
--hf-repo Abhiray/Gemma-4-12B-it-AEON-Abliterated-K4-GGUF \
--hf-file Gemma-4-12B-it-AEON-Abliterated-Q4_K_M.gguf \
-ngl -1 \
-c 4096 \
-p "<start_of_turn>user\nYour prompt here<end_of_turn>\n<start_of_turn>model\n"