llmfan46/gemma-3-12b-it-ultra-uncensored-heretic

VISIONPricing:Input $0.2 / Output $0.6Concurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kPublished:Mar 9, 2026License:gemmaArchitecture:Transformer0.0K Featherless Exclusive Cold

llmfan46/gemma-3-12b-it-ultra-uncensored-heretic is a 12 billion parameter instruction-tuned multimodal model, derived from Google's Gemma 3, with a 32768 token context window. This version has been decensored using the Heretic v1.2.0 tool with Arbitrary-Rank Ablation (ARA) to significantly reduce content refusals from 97/100 to 4/100, while maintaining original model capabilities with a low KL divergence of 0.0131. It is designed for applications requiring less restrictive content generation, supporting text and image inputs for tasks like question answering, summarization, and reasoning.

Loading preview...

Model Overview

This model, llmfan46/gemma-3-12b-it-ultra-uncensored-heretic, is a 12 billion parameter instruction-tuned variant of Google's Gemma 3, enhanced for reduced content restrictions. It leverages the Heretic v1.2.0 tool with the Arbitrary-Rank Ablation (ARA) method to achieve its 'decensored' state. The original Gemma 3 model is a multimodal architecture capable of processing both text and image inputs (normalized to 896x896 resolution) and generating text outputs, supporting a 32K token context window for this size.

Key Differentiators

  • Decensored Behavior: The primary distinction is its significantly reduced refusal rate, dropping from 97/100 in the original gemma-3-12b-it to just 4/100. This makes it suitable for use cases where the original model's content restrictions were prohibitive.
  • Capability Preservation: Despite decensoring, the model maintains a low KL divergence of 0.0131, indicating strong preservation of the original Gemma 3's coherence, reasoning ability, and overall quality.
  • Multimodal Input: Inherits Gemma 3's ability to handle both text and image inputs, making it versatile for various applications requiring visual understanding alongside text generation.
  • Multilingual Support: Based on Gemma 3, it supports over 140 languages.

Intended Use Cases

This model is well-suited for developers and researchers who require a powerful, multimodal language model with fewer built-in content restrictions. It can be applied to:

  • Creative Content Generation: For scenarios where more freedom in expression is desired.
  • Conversational AI: Building chatbots or virtual assistants that require less filtering on user inputs or generated responses.
  • Research and Experimentation: Exploring model behaviors and capabilities without the constraints of strict safety filters.
  • Text and Image Understanding: Tasks like question answering, summarization, and reasoning from combined text and image inputs, with a focus on unrestricted output.