llmfan46/gemma-4-26B-A4B-it-ultra-uncensored-heretic
The llmfan46/gemma-4-26B-A4B-it-ultra-uncensored-heretic is a 26 billion parameter instruction-tuned causal language model, a decensored version of Google's Gemma-4-26B-A4B-it. Developed by llmfan46 using the Heretic tool with Arbitrary-Rank Ablation (ARA) method, it significantly reduces content refusals by 97% while largely preserving original model quality. This model is optimized for applications requiring less restrictive content generation, maintaining strong performance in reasoning, coding, and multimodal understanding.
Loading preview...
Model Overview
This model, llmfan46/gemma-4-26B-A4B-it-ultra-uncensored-heretic, is a 26 billion parameter instruction-tuned variant of Google's Gemma-4-26B-A4B-it. It has been decensored using the Heretic tool with the Arbitrary-Rank Ablation (ARA) method, specifically targeting attn.o_proj components across layers 11 to 22.
Key Differentiators & Performance
- Reduced Refusals: Achieves a remarkable 97% reduction in content refusals (3/100 vs. 100/100 for the original model), making it suitable for less restricted content generation.
- Preserved Quality: Maintains most of the original model's quality with a low KL divergence of 0.1237.
- Reasoning & Multimodality: Inherits Gemma 4's core capabilities, including strong reasoning, coding, and multimodal understanding (text, image, video).
- Benchmarks: While significantly reducing refusals, it shows a slight decrease in MMLU accuracy (79.99% vs. 82.48% for original) and PIQA accuracy (91.29% vs. 92.06% for original), indicating a trade-off for uncensored behavior.
Good for
- Applications requiring a less restrictive content policy.
- Tasks involving text generation, coding, and reasoning where some quality trade-off for uncensored output is acceptable.
- Developers seeking to experiment with decensored large language models based on the Gemma 4 architecture.