DavidAU/gemma-3-12b-it-vl-Minimax-M2.1-Heretic-Uncensored-Thinking
DavidAU's gemma-3-12b-it-vl-Minimax-M2.1-Heretic-Uncensored-Thinking is a 12 billion parameter Gemma 3 fine-tune, developed by DavidAU, featuring a 32768 token context length. This model is explicitly uncensored and optimized for deep reasoning across general operation, output generation, and image processing, utilizing the Minimax-M2.1 reasoning dataset. It is designed to provide direct, detailed responses without refusal, making it suitable for use cases requiring explicit or nuanced content generation.
Loading preview...
Model Overview
DavidAU/gemma-3-12b-it-vl-Minimax-M2.1-Heretic-Uncensored-Thinking is a 12 billion parameter Gemma 3 instruction-tuned model, fine-tuned by DavidAU using the Minimax-M2.1 reasoning dataset. This model is notable for being fully uncensored and engineered for deep reasoning, which enhances its general operation, output generation, and image processing capabilities. It offers a substantial 32768 token context length and maintains reasoning stability across a wide temperature range (.1 to 2.5).
Key Capabilities
- Uncensored Output: Designed to generate content without refusal, including explicit or sensitive topics, though it may require specific directives for desired intensity.
- Enhanced Reasoning: Integrates deep thinking logic, improving the quality and detail of responses across various tasks.
- Image Processing: Reasoning capabilities extend to image processing tasks.
- Flexible Activation: Reasoning activates automatically but can be explicitly triggered with "think deeply: prompt" or via specialized Jinja templates and system prompts.
Performance & Uncensoring
Benchmarks indicate competitive performance against its uncensored base, with specific scores provided for arc_challenge, arc_easy, boolq, hellaswag, openbookqa, piqa, and winogrande. The model demonstrates a KL divergence of 0.0826 compared to the original google/gemma-3-12b-it, indicating minimal damage from the uncensoring process, and significantly reduced refusals (7/100 vs. 98/100).
Optimal Usage
For best results, users can leverage optional system prompts to guide reasoning or utilize the chat-template-thinking.jinja for always-on thinking. Settings like Smoothing_factor (1.5) in interfaces like KoboldCpp or text-generation-webui are recommended for smoother operation, especially for chat and roleplay scenarios.