TrevorJS/gemma-4-12B-it-uncensored

TEXT GENERATIONConcurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 13, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

TrevorJS/gemma-4-12B-it-uncensored is a 12 billion parameter instruction-tuned causal language model based on Google's Gemma-4 architecture, specifically the encoder-free Gemma4Unified model. This version has been modified to significantly reduce refusal behavior, achieving an effective refusal rate of ~0/686 on cross-dataset validation, while maintaining response quality. It is optimized for use cases requiring less restrictive content generation without degradation in coherence.

Loading preview...

Overview

TrevorJS/gemma-4-12B-it-uncensored is a modified version of Google's gemma-4-12B-it model, specifically engineered to remove refusal behaviors. This 12 billion parameter model, based on the Gemma4Unified architecture, has undergone a process called norm-preserving biprojected abliteration to achieve this. The modification targets refusal signals concentrated in the upper decoder layers (L15-47), ablating 70% of these layers to remove refusal tendencies without introducing distortion.

Key Capabilities

  • Reduced Refusal Behavior: Achieves a significant reduction in refusals, from 99/100 to 6/100 on mlabonne prompts, and an effective rate of ~0/686 across multiple independent datasets (JailbreakBench, tulu-harmbench, NousResearch/RefusalDataset, mlabonne/harmful_behaviors).
  • Quality Preservation: Manual audits and Q8 inference verified that the removal of refusal behavior did not degrade response quality or coherence.
  • Norm-Preserving Modification: Utilizes a unique method that preserves weight magnitudes, ensuring model stability and performance.
  • Efficient Abliteration: Employs per-layer refusal directions and a deterministic single-pass process, offering efficiency over traditional methods.

Good For

  • Applications requiring a less restrictive language model for content generation.
  • Use cases where the original Gemma-4 model's refusal mechanisms were overly cautious or hindered desired outputs.
  • Developers seeking a powerful 12B parameter model with enhanced flexibility in response generation.

Finetunes
2 models