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

TEXT GENERATIONPricing:Input $1.2 / Cached $0.24 / Output $4.8Concurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 14, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Iambackup/gemma-4-12B-it-uncensored is a 12 billion parameter instruction-tuned 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 a refusal rate of approximately 0/686 on cross-dataset validation, while maintaining response quality. It utilizes a norm-preserving biprojected obliteration method to remove refusal signals from the model weights, making it suitable for applications requiring less constrained output.

Loading preview...

Overview

Iambackup/gemma-4-12B-it-uncensored is an instruction-tuned 12 billion parameter model derived from Google's gemma-4-12B-it. Its primary distinction is the removal of refusal behavior, making it an "uncensored" variant. This modification was achieved using a novel norm-preserving biprojected obliteration method, which selectively projects out refusal directions from the model's weights while preserving their magnitudes. The process targets the upper decoder layers (L15-47) of the Gemma4Unified architecture, where refusal signals are concentrated.

Key Capabilities

  • Significantly Reduced Refusal Behavior: Achieves an effective refusal rate of approximately 0/686 across multiple independent prompt datasets, down from 99/100 in the original model on specific tests.
  • Maintained Response Quality: Manual audits and Q8 inference verified no degradation in response coherence or quality after the uncensoring process.
  • Norm-Preserving Modification: Employs a method that ensures ||W_new|| = ||W_orig||, preventing distortion of the model's original weight magnitudes.
  • Efficient Abliteration: Utilizes a deterministic single-pass method for faster and equally effective refusal removal compared to multi-trial search methods.

Good for

  • Applications requiring a language model with minimal content restrictions or refusal to answer.
  • Research into model safety, bias, and methods for controlling model behavior.
  • Use cases where the original gemma-4-12B-it's refusal tendencies were a limiting factor.
  • Developers needing a 12B parameter model with a 32768 token context length that offers more direct and less filtered responses.

Finetunes
2 models