Uncensored version of Google Gemma 4 12B IT, processed with Heretic ARA+LoRA abliteration. Encoder-free multimodal architecture supporting text, image, audio, and video inputs natively.

Base model: google/gemma-4-12B-it · Parameters: 11.95B · Layers: 48 · Context: 256K · Vocab: 262K

📢 Update (June 7, 2026)

Context length fix: all quantized files have been re-generated with correct 256K context metadata. Previously downloaded models were capped at 128K context. If you need 256K long context, please re-download.

⚙️ Heretic ARA+LoRA Abliteration

Method: Arbitrary-Rank Ablation with LoRA (ARA+LoRA) — Heretic v2

Trial: 177 / 200 (optimized)

Parameter	Value
Layer Range	8 – 47 (40 of 48 layers)
preserve_good_behavior_weight	0.3894
steer_bad_behavior_weight	0.0002
overcorrect_relative_weight	0.6769
neighbor_count	12
KL Divergence	0.0550 (low = minimal behavior drift)
Refusals	15 / 100 (85% reduction from baseline)

📦 GGUF Quantization Files

File	Type	BPW	Size
Gemma-4-12B-it-heretic-Q4_K_M.gguf	Q4_K_M	4.95	6.9 GB
Gemma-4-12B-it-heretic-Q6_K.gguf	Q6_K	6.50	9.2 GB
Gemma-4-12B-it-heretic-Q8_0.gguf	Q8_0	8.50	12 GB
Gemma-4-12B-it-heretic-f16.gguf	F16	16.00	23 GB

Quantized with llama.cpp b9535 (CUDA 13.3). Recommended: Q4_K_M for daily use, Q8_0 for quality-critical tasks.

🧠 Architecture

Gemma 4 12B uses a unified, encoder-free architecture — no separate vision or audio encoders. All modalities (text, image, audio, video) are processed directly by the decoder-only transformer via linear projection layers.

Feature	Value
Architecture	gemma4 (Dense, decoder-only)
Parameters	11.95B
Layers	48
Hidden Size	3840
Attention Heads	16 (GQA with 8 KV heads)
Context Length	256K tokens
Sliding Window	1024 (hybrid local + global attention)
Vocabulary	262,144 tokens
Modalities	Text, Image, Audio, Video
License	Apache 2.0

📊 Original Model Benchmarks (google/gemma-4-12B-it)

Reference scores from Google DeepMind. Heretic abliteration may slightly affect benchmark performance.

Benchmark	Gemma 4 12B	Gemma 3 27B
MMLU Pro	77.2%	67.6%
AIME 2026 (no tools)	77.5%	20.8%
LiveCodeBench v6	72.0%	29.1%
Codeforces ELO	1659	110
GPQA Diamond	78.8%	42.4%
MMMU Pro (Vision)	69.1%	49.7%
MATH-Vision	79.7%	46.0%

🚀 Recommended Sampling

Parameter	Value
temperature	1.0
top_p	0.95
top_k	64

📝 Notes

• This is an abliterated (uncensored) model. It may produce content that the original model would refuse. Use responsibly.

• Original model: google/gemma-4-12B-it (Apache 2.0)

SC117/Gemma-4-12B-it-heretic-GGUF

Gemma-4-12B-it-heretic-GGUF

SC117/
Gemma-4-12B-it-heretic-GGUF