LastRef/gemma-3-12b-it-heretic-x: Decensored Multimodal Gemma 3

This model is a 12 billion parameter instruction-tuned variant of Google's Gemma 3, specifically engineered by LastRef as a decensored version using the Heretic tool. Unlike typical Heretic models, it was developed with a custom x-rated dataset and fine-tuned for less restrictive content generation. It maintains the multimodal capabilities of the base Gemma 3, handling both text and image inputs to generate text outputs, with a 32K token context window.

Key Capabilities

• Decensored Content Generation: Optimized for producing less restricted and x-rated content, outperforming other Gemma-3-12b-it ablations in internal NSFW evaluations.
• Multimodal Input: Processes both text and image inputs (images normalized to 896x896 resolution, encoded to 256 tokens each).
• Strong Writing Performance: Achieves the 2nd place in writing categories for Gemma-3-12b models on the UGI-Leaderboard.
• Broad Language Support: Inherits multilingual support in over 140 languages from the base Gemma 3.

Good For

• Applications requiring decensored or x-rated text generation.
• Creative writing and roleplay scenarios with fewer content restrictions.
• Research and experimentation into model safety and content filtering bypasses.
• Multimodal tasks such as image analysis combined with unrestricted text generation.