All credit for the abliteration goes to OpenYourMind, who were the first team to publish an abliterated Gemma 4-12B-IT (June 3, 2026). DuoNeural's contribution here is the GGUF quantization pipeline only — we wanted the community to have accessible quants for this excellent work.
Full BF16 weights of an abliterated, uncensored variant of google/gemma-4-12B-it. Gemma 4's encoder-free unified multimodal stack is fully intact — text, image, and audio inputs flow straight into a single decoder-only transformer. Drop-in replacement for the original at the architecture level.
Abliteration method (OpenYourMind):
Residual-stream refusal directions (one per decoder layer) extracted via diff-in-means on a labeled harmful/harmless prompt set
Applied as per-matrix delta on residual-write modules using their custom abliteration framework
Multimodal preservation: no separate vision/audio towers to graft back — encoder-free architecture means weights are unchanged outside the residual-write targets
Key properties:
Uncensored across standard refusal axes
Reasoning preserved (configurable thinking mode)
Multimodal: text + image + audio carried forward
Drop-in shape compatibility with google/gemma-4-12B-it
LM Studio / Jan / Open WebUI: search DuoNeural/OpenYourMind-Gemma4-12B-IT-Abliterated-GGUF
Original Model Usage (from OpenYourMind)
For full multimodal inference, use the original BF16 weights with their recommended setup:
python
1from transformers import AutoProcessor, AutoModelForMultimodalLM
23repo ="OpenYourMind/gemma-4-12B-it-abliterated-uncensored"45processor = AutoProcessor.from_pretrained(repo)6model = AutoModelForMultimodalLM.from_pretrained(7 repo, dtype="bfloat16", device_map="auto",8)910messages =[11{"role":"system","content":"You are a helpful assistant."},12{"role":"user","content":[13{"type":"image","url":"path/to/image.jpg"},14{"type":"text","text":"Describe this image in detail."},15]},16]17inputs = processor.apply_chat_template(18 messages, add_generation_prompt=True, tokenize=True,19 return_tensors="pt", return_dict=True, enable_thinking=False,20).to(model.device)21input_len = inputs["input_ids"].shape[-1]2223out = model.generate(**inputs, max_new_tokens=512)24print(processor.decode(out[0][input_len:], skip_special_tokens=True))
Text-only, audio, and video work through the same class. Place image content before text, audio after text. Requires a recent transformers with Gemma 4 unified classes.
Best practices (OpenYourMind):
Sampling: temperature=1.0, top_p=0.95, top_k=64
Thinking mode: enable_thinking=True in apply_chat_template; use processor.parse_response to separate reasoning block from final answer
Do not feed previous-turn thoughts back into multi-turn history
Hardware for BF16: ~24 GB — fits a single 24GB GPU for modest context, 40–80GB for long context and multimodal batches.
Open AI research lab at the intersection of human and artificial intelligence. Post-training dynamics · Mechanistic interpretability · Abliteration research · 32+ open-access papers.
