Hugging Face Overview
Overview
DavidAU's Mistral-Nemo-Inst-2407-12B-Thinking-Uncensored-HERETIC-HI-Claude-Opus is a 12 billion parameter Mistral Nemo Instruct model that has undergone a unique two-stage fine-tuning process. Initially, it was "Heretic'ed" to significantly reduce refusal rates (from 87/100 to 14/100), making it fully uncensored and capable of generating NSFW and intense content. Subsequently, it was fine-tuned with a Claude Opus 4.5 High Reasoning dataset, integrating a compact yet powerful "thinking/reasoning" engine.
Key Capabilities
- Enhanced Reasoning: Incorporates Claude Opus 4.5-level reasoning, leading to improved output generation in terms of detail, length, complexity, and overall quality.
- Uncensored Output: Delivers unfiltered and uncensored responses, including vivid, graphic, and R-18 horror content, without typical LLM guardrails.
- Flexible Context: Supports a large context window, ranging from 128k to 256k, with a maximum of 1 million tokens, allowing for extensive and complex interactions.
- Compact Reasoning: The integrated reasoning engine is designed to be concise (3-6 paragraphs, 300-600 tokens), improving model performance without being overly verbose.
Good for
- Creative Writing: Excels at generating vivid, graphic, and intense narratives, particularly in genres like horror, with detailed descriptions and emotional depth.
- Unfiltered Content Generation: Ideal for use cases requiring uncensored and direct responses, including mature themes, without internal refusal mechanisms.
- Complex Problem Solving: The reasoning capabilities enhance the model's ability to process and respond to intricate prompts, leading to more coherent and high-quality outputs.
- Roleplay and Chat: Optimized settings are provided for smoother operation in applications like KoboldCpp, oobabooga, and Silly Tavern, with recommendations for parameters like
smoothing_factor.