Uncensored version of Google Gemma 4 26B-A4B 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.

✂️ Heretic ARA Abliteration Parameters

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

Parameter	Value
start_layer_index	12
end_layer_index	21
preserve_good_behavior_weight	0.3106
steer_bad_behavior_weight	0.0066
overcorrect_relative_weight	0.7982
neighbor_count	14

Metric	Heretic	Original QAT
KL Divergence	0.0660	0 (by definition)
Refusals	13/100	100/100

🏗️ Architecture

Base Model	google/gemma-4-26B-A4B-it
Parameters	25.2B total / 3.8B active (MoE)
Architecture	Mixture-of-Experts: 128 experts, 8 active + 1 shared per token
Layers	30
Hidden Size	2,816
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 (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	14 GB
Effective BPW	4.51 (all weight tensors Q4_0, norms/router 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 85.6% Top-1 vs 70.2% naive Q4_0 (+15.4%)

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. Use --jinja flag and -ngl 99 for GPU offload. Vision support via included mmproj.

llama-server \ -m gemma-4-26B-A4B-it-qat-heretic-UD-Q4_K_XL.gguf \ --mmproj mmproj-gemma-4-26B-A4B-it-qat-heretic-BF16.gguf \ --jinja -ngl 99 -c 8192 \ --port 8001

🔗 Credits

Heretic Abliteration: coder3101 · Heretic v1.2.0 ARA + Row-Norm
QAT Weights: Google Gemma 4 QAT
Quantization Recipe: Unsloth UD-Q4_K_XL (Q4_0 layout)
Quantization Tool: llama.cpp b9553 · GitHub
Original Model: Google Gemma 4 26B-A4B IT

SC117/gemma-4-26B-A4B-it-qat-heretic-GGUF

⚡ Gemma 4 26B-A4B Heretic QAT — Q4_0 GGUF

SC117/
gemma-4-26B-A4B-it-qat-heretic-GGUF