Ishowbackup/gemma-4-26B-A4B-it-uncensored

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:2Model Size:26BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 18, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Ishowbackup/gemma-4-26B-A4B-it-uncensored is a 26 billion parameter instruction-tuned causal language model based on Google's Gemma-4 architecture. This model has been specifically modified to significantly reduce refusal behavior, achieving a 0.7% refusal rate across multiple datasets, while maintaining original response quality. It is optimized for use cases requiring direct answers without AI identity disclaimers or content filtering.

Loading preview...

Model Overview

This model, Ishowbackup/gemma-4-26B-A4B-it-uncensored, is an uncensored variant of the google/gemma-4-26B-A4B-it model. It has been meticulously modified to remove refusal behavior, making it suitable for applications where direct and unfiltered responses are preferred. The modification process involved advanced abliteration techniques, including Expert-Granular Abliteration (EGA), applied to the model's dense pathways and MoE expert weights.

Key Capabilities

  • Significantly Reduced Refusals: Achieves a refusal rate of only 0.7% across 686 prompts from various datasets (JailbreakBench, tulu-harmbench, NousResearch/RefusalDataset, mlabonne/harmful_behaviors).
  • Quality Preservation: Maintains the original response quality, with a harmless response length ratio of approximately 1.01, indicating no degradation in output quality.
  • Advanced Abliteration Method: Utilizes norm-preserving biprojected abliteration on o_proj and mlp.down_proj, combined with EGA on all 128 MoE expert down_proj slices per layer, a technique not supported by vanilla heretic methods.
  • Efficient Processing: Employs a deterministic single-pass process for refusal direction computation, offering faster and equally effective results compared to multi-trial search methods.

Good For

  • Applications requiring direct, unfiltered responses without AI identity disclaimers or refusal to answer.
  • Research into model safety and bias removal techniques.
  • Use cases where the base Gemma-4 model's refusal behavior is a limitation.

Finetunes
2 models