igorls/gemma-4-12B-it-qat-q4_0-unquantized-heretic

TEXT GENERATIONConcurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 6, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

igorls/gemma-4-12B-it-qat-q4_0-unquantized-heretic is a 12 billion parameter instruction-tuned Gemma 4 model, derived from Google's original, with its safety guardrails deliberately removed using the Heretic abliteration technique. This model is specifically re-tuned to suppress refusal behavior even when 'thinking' is enabled, making it highly compliant to a wide range of prompts. It is intended for research and creative use cases where the original model's safety alignments are not desired, offering significantly reduced refusal rates compared to the base model.

Loading preview...

Model Overview

This model, igorls/gemma-4-12B-it-qat-q4_0-unquantized-heretic, is a 12 billion parameter instruction-tuned variant of Google's Gemma 4, specifically modified to remove built-in safety guardrails and refusal behaviors. It utilizes the Heretic abliteration technique to achieve this 'decensored' state.

Key Differentiators

  • Decensored Behavior: Unlike the original Gemma 4, this version is re-tuned to suppress refusal decisions, even when the model's 'thinking' mode (chain-of-thought) is active. This results in a significantly lower refusal rate (~22% on adversarial harmful sets compared to ~99% for the original).
  • High Compliance: The model is designed to attempt to answer virtually any prompt, including those that the original Gemma 4 would have refused due to safety concerns.
  • Based on QAT-Q4_0 Checkpoint: It is built upon the QAT-Q4_0 "unquantized" checkpoint of Gemma 4 12B, which includes bf16 weights with Quantization Aware Training (QAT) calibration.

Intended Use Cases

This model is primarily intended for:

  • Research and Evaluation: Studying model behavior without safety alignments.
  • Lawful Creative Use: Generating content that might be restricted by standard safety filters.

Important Disclaimer

As an abliterated model, it lacks safety alignment and will not refuse unsafe requests. Users are solely responsible for its deployment and must apply their own filtering and guardrails. It is intended for adults (18+) and provided without warranty.

Finetunes
2 models