gemma-4-12b-it-heretic (Q4)

n0404n0404/gemma-4-12b-it-heretic-b9a462:latest

A decensored build of Google’s Gemma 4 12B IT, processed with Heretic, quantized to Q4, and published on Ollama.

Model Details

Quick Start

# Pull the model
ollama pull n0404n0404/gemma-4-12b-it-heretic-b9a462:latest

# Start chatting
ollama run n0404n0404/gemma-4-12b-it-heretic-b9a462:latest

Via the API:

curl http://localhost:11434/api/chat -d '{
  "model": "n0404n0404/gemma-4-12b-it-heretic-b9a462:latest",
  "messages": [
    { "role": "user", "content": "Hello, introduce yourself." }
  ]
}'

What Is This?

Heretic (decensoring)

Heretic is an automated decensoring tool for language models. It uses directional ablation (also known as abliteration) to identify and remove the internal directions responsible for refusal behavior, which lowers the model’s rate of unnecessary refusals.

Its main selling point is that the process is fully automatic: it optimizes the ablation parameters to balance a low refusal rate against preservation of the original model’s capabilities (measured as KL divergence from the base model). No manual tuning and no additional training data are required.

Q4 quantization

After decensoring, the model is quantized to Q4 (4-bit) GGUF. This substantially reduces memory and VRAM footprint, letting a 12B model run on consumer hardware, at the cost of a small quality loss compared to full precision.

Pipeline

google/gemma-4-12B-it   (base model)
        │
        ▼
   Heretic              (directional ablation / refusal removal)
        │
        ▼
   Q4 quantization      (4-bit GGUF)
        │
        ▼
   Published to Ollama

Hardware Requirements (approximate)

Actual usage varies with context length and runtime configuration.

⚠️ Important Notes

• This model has been decensored, so its safety guardrails are significantly weakened. It may produce unfiltered, offensive, inaccurate, or otherwise inappropriate content. Use your own judgment and take responsibility for how you use its output.
• Decensoring and quantization both alter model behavior. The outputs of this model do not represent the positions or quality of Google or the original Gemma model.
• Do not use this model for anything illegal, to harm others, or to generate malicious content.
• Outputs may contain factual errors. Verify anything important yourself before relying on it.

License

• The base Gemma model is distributed under the Gemma Terms of Use. Use of this derivative model is subject to those terms and to the Gemma Prohibited Use Policy.
• The decensored and quantized derivatives are bound by the same terms.

Credits

• Google Gemma — base model
• Heretic by p-e-w — decensoring tool
• Ollama — model runtime and registry