llmfan46/Gemma-4-Garnet-31B-it-uncensored-heretic

VISIONPricing:Input $0.48 / Cached $0.1 / Output $1.44Concurrent Unit Cost:2Model Size:31BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Apr 14, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Gemma-4-Garnet-31B-it-uncensored-heretic by llmfan46 is a 31 billion parameter instruction-tuned language model, a decensored version of ConicCat/Gemma4-Garnet-31B. It significantly reduces content refusals by 94% (6/100 vs 99/100) while maintaining model quality with a KL divergence of 0.0368. This model is primarily optimized for creative writing and translation, offering improved longform writing capabilities, reduced 'AI slop', and enhanced output diversity, making it suitable for applications requiring less restrictive and more nuanced text generation.

Loading preview...

Overview

llmfan46/Gemma-4-Garnet-31B-it-uncensored-heretic is a 31 billion parameter instruction-tuned model, derived from ConicCat/Gemma4-Garnet-31B. This version has been decensored using the Heretic v1.2.0 tool with the Arbitrary-Rank Ablation (ARA) method, specifically targeting attn.o_proj components across layers 26-46.

Key Capabilities

  • Significantly Reduced Refusals: Achieves a 94% reduction in content refusals (6/100 compared to 99/100 in the original model), allowing for less restricted and more flexible text generation.
  • Preserved Model Quality: Maintains a low KL divergence of 0.0368 from the original model, indicating that the decensoring process has minimal impact on baseline performance.
  • Improved Writing Quality: Enhances creative writing and translation capabilities, addressing stiffness and odd word choices present in the base model.
  • Longform Generation: Features improved longform writing, capable of generating up to 4000 words of text in a single output.
  • Reduced 'AI Slop': Produces text with markedly less 'AI slop' and identifiable Gemini-isms, leading to more natural-sounding prose.
  • Enhanced Output Diversity: Offers improved output diversity, reducing 'soft' refusals and increasing the variety of generated content.

Performance Highlights

  • PIQA Accuracy: Shows a slight improvement in PIQA (Physical Intuition Question Answering) accuracy, from 93.63% (original) to 93.80% (Heretic).
  • MMLU Performance: Exhibits a minor decrease in MMLU (Massive Multitask Language Understanding) accuracy, from 85.91% (original) to 84.80% (Heretic), which is a trade-off for the significant reduction in refusals.

Good For

  • Creative Writing: Ideal for generating diverse and less restrictive creative content, including roleplay scenarios and longform narratives.
  • Translation: Suitable for translation tasks where nuanced and less 'stiff' language is preferred.
  • Applications Requiring Uncensored Output: Excellent for use cases where content restrictions are undesirable, such as open-ended conversational agents or content generation platforms that require a broader range of expression.

Finetunes
2 models