--- title: gemma-4-26B-A4B-it-qat-heretic-GGUF canonical_url: "https://www.modelscope.ai/models/SC117/gemma-4-26B-A4B-it-qat-heretic-GGUF" md_url: "https://www.modelscope.ai/models/SC117/gemma-4-26B-A4B-it-qat-heretic-GGUF.md" repository: SC117/gemma-4-26B-A4B-it-qat-heretic-GGUF last_updated: 2026-10-04 license: apache-2.0 base_model: - coder3101/gemma-4-26B-A4B-it-qat-q4_0-unquantized-heretic - google/gemma-4-26B-A4B-it-qat-q4_0-unquantized base_model_relation: quantized library_name: - gguf - pytorch frameworks: - pytorch language: - en - zh - multilingual downloads: 138 stars: 0 tags: - gguf - quantized - qat - heretic - uncensored - abliterated - moe - gemma4 --- # gemma-4-26B-A4B-it-qat-heretic-GGUF > gemma-4-26B-A4B-it-qat-heretic-GGUF - An open-source model by SC117 on ModelScope. ⚡ Gemma 4 26B-A4B Heretic QAT — Q40 GGUF Heretic ARA · QAT-Lossless Q40 · 14 GB · MoE 128-Expert (3.8B Active) 📖 中文文档 Q40 26B MoE / 3.8B Active Heretic Uncensored 14 GB QAT… SC117/gemma-4-26B-A4B-it-qat-heretic-GGUF is a machine learning model on ModelScope. licensed under apache-2.0. derived from coder3101/gemma-4-26B-A4B-it-qat-q4_0-unquantized-heretic、google/gemma-4-26B-A4B-it-qat-q4_0-unquantized. - **Repository**: SC117/gemma-4-26B-A4B-it-qat-heretic-GGUF - **License**: apache-2.0 - **Base model**: coder3101/gemma-4-26B-A4B-it-qat-q4_0-unquantized-heretic, google/gemma-4-26B-A4B-it-qat-q4_0-unquantized - **Tags**: gguf, quantized, qat, heretic, uncensored, abliterated, moe, gemma4 - **Downloads**: 138 - **Stars**: 0 - **Last updated**: 2026-10-04 Source: https://www.modelscope.ai/models/SC117/gemma-4-26B-A4B-it-qat-heretic-GGUF ---

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

Heretic ARA · QAT-Lossless Q4_0 · 14 GB · MoE 128-Expert (3.8B Active)

📖 中文文档

Q4_0 26B MoE / 3.8B Active Heretic Uncensored 14 GB QAT Weights 128 Experts

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_index12
end_layer_index21
preserve_good_behavior_weight0.3106
steer_bad_behavior_weight0.0066
overcorrect_relative_weight0.7982
neighbor_count14
Metric Heretic Original QAT
KL Divergence0.06600 (by definition)
Refusals13/100100/100
🏗️ Architecture
Base Modelgoogle/gemma-4-26B-A4B-it
Parameters25.2B total / 3.8B active (MoE)
ArchitectureMixture-of-Experts: 128 experts, 8 active + 1 shared per token
Layers30
Hidden Size2,816
Attention16 heads, GQA with 8 KV heads, head dim 256
Context Length256K tokens (hybrid sliding window 1024 + global attention)
Vocabulary262K, 140+ languages
ModalitiesText + Image (native multimodal)
QAT TrainingGoogle official QAT (quantization-aware), weights inherently robust to Q4_0
QuantizationQ4_0 (matching Unsloth UD-Q4_K_XL layout), b9553 llama-quantize
📊 Quantization Details
FormatQ4_0 (uniform — QAT weights optimized for this exact precision)
File Size14 GB
Effective BPW4.51 (all weight tensors Q4_0, norms/router F32)
Toolllama-quantize (b9553, CUDA 13.3)
SourceBF16 GGUF (converted from QAT heretic safetensors)
Context Length256K (set in GGUF metadata)
QAT AdvantageQ4_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
Generaltemp=1.0, top_p=0.95, top_k=64
Codingtemp=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

Support SC117 on Ko-fi — support is optional and all releases remain freely available.