Uncensored version of Google Gemma 4 12B IT (QAT), processed with Heretic ARA abliteration. Quantized to Q4_0 matching Unsloth's UD-Q4_K_XL format — QAT weights trained for 4-bit quantization, near-lossless quality.

⚠️ Important Notice: UD-Q4_K_XL and Google QAT

UD-Q4_K_XL is the UD team's naming convention for its Gemma QAT Q4_0 GGUF variant. The repository or file name must not be interpreted as proof that the model uses mixed K-quants. In this release, the actual GGUF tensor information shows that the main model weights are stored as Q4_0, while normalization and scaling tensors remain in F32.

This model is derived from Google's QAT weights, which were trained for 4-bit quantization. Claims that this file was converted to UD mixed precision, or that its QAT structure was damaged by mixed K-quant remapping, are therefore not supported by the actual tensor types contained in the GGUF.

Before opening a discussion or reporting a quantization issue, please inspect the GGUF metadata and tensor types first. Do not infer the internal quantization layout from the filename alone, and do not make unsupported technical claims without checking the source information.

✂️ Heretic ARA Abliteration Parameters

Base: coder3101/heretic-QAT · Heretic v1.2.0 · ARA + Row-Norm

Parameter	Value
start_layer_index	24
end_layer_index	48
preserve_good_behavior_weight	0.3707
steer_bad_behavior_weight	0.0010
overcorrect_relative_weight	0.6177
neighbor_count	15

Metric	Heretic	Original QAT
KL Divergence	0.0575	0 (by definition)
Refusals	8/100	99/100

🏗️ Architecture

Base Model	google/gemma-4-12B-it
Parameters	11.95B (dense, all parameters active)
Architecture	Encoder-free unified multimodal (text + image + audio + video)
Layers	48
Hidden Size	3,840
Attention	16 heads, GQA with 8 KV heads, head dim 256
Context Length	256K tokens (hybrid sliding window 1024 + global attention)
Vocabulary	262K, 140+ languages
Modalities	Text + Image + Audio + Video (encoder-free, native multimodal)
QAT Training	Google official QAT (quantization-aware), weights inherently robust to Q4_0
Quantization	Q4_0 (matching Unsloth UD-Q4_K_XL layout), b9553 llama-quantize

📊 Quantization Details

Format	Q4_0 (uniform — QAT weights optimized for this exact precision)
File Size	6.4 GB
Effective BPW	4.50 (all weight tensors Q4_0, norms F32)
Tool	llama-quantize (b9553, CUDA 13.3)
Source	BF16 GGUF (converted from QAT heretic safetensors)
Context Length	256K (set in GGUF metadata)
QAT Advantage	Q4_0 with QAT weights achieves 88.8% Top-1 vs 74.1% naive Q4_0 (+14.7%)

Why Q4_0? Google's QAT trains weights to be optimal at Q4_0 noise levels. Unsloth's UD-Q4_K_XL uses the same Q4_0 layout — the "dynamic" advantage comes from conversion precision, not per-tensor mixing.

⚙️ Recommended Sampling Parameters

General	temp=1.0, top_p=0.95, top_k=64
Coding	temp=0.6, top_p=0.95, top_k=64

Use --jinja flag with llama.cpp. Disable thinking: --chat-template-kwargs '{"enable_thinking":false}'.

📝 Usage

Compatible with llama.cpp, LM Studio, Jan, koboldcpp, and other GGUF runtimes. Fits easily on 8GB VRAM. Encoder-free — no separate vision/audio projector needed.

llama-server \ -m gemma-4-12B-it-heretic-QAT-UD-Q4_K_XL.gguf \ --jinja -ngl 99 -c 8192 \ --port 8001

🔗 Credits

Heretic Abliteration: coder3101 · Heretic v1.2.0 ARA + Row-Norm
QAT Weights: Google Gemma 4 12B IT
Quantization Recipe: Unsloth UD-Q4_K_XL (Q4_0 layout)
Quantization Tool: llama.cpp b9553 · GitHub
Original Model: Google Gemma 4 12B IT

SC117/gemma-4-12B-it-heretic-QAT-GGUF

⚡ Gemma 4 12B Heretic QAT — Q4_0 GGUF

SC117/
gemma-4-12B-it-heretic-QAT-GGUF