Overview

DavidAU/gemma-3-12b-it-vl-Deepseek-v3.1-Heretic-Uncensored-Thinking is a 12 billion parameter Gemma fine-tune, developed by DavidAU, focusing on uncensored and deep reasoning. It utilizes the Deepseek 3.1 reasoning dataset and is trained via Unsloth. The model features a 128k context window and stable reasoning across a temperature range of 0.1 to 2.5.

Key Capabilities

• Uncensored Output: Designed to provide direct responses without refusal, even for sensitive content, though it may require explicit direction for graphic or explicit levels.
• Enhanced Reasoning: Improves general model operation, output generation, image processing, and benchmark performance. Reasoning can be activated explicitly with "think deeply: prompt" or through optional system prompts and a specialized Jinja template.
• High KL Divergence Score: Achieves a KL divergence of 0.0826, indicating minimal damage to the original model's distribution during de-censoring, with significantly reduced refusals (7/100 compared to 98/100 for the original).

Good For

• Applications requiring unrestricted and direct content generation.
• Use cases where deep and detailed reasoning is critical for accurate and precise outputs.
• Scenarios benefiting from enhanced image processing through advanced reasoning.
• Developers seeking a model with flexible control over output style through system prompts and thinking activation mechanisms.