license: apache-2.0
base_model: MuXodious/Ministral-3-3B-Instruct-2512-absolute-heresy
tags:
- unfiltered
- visceral
- storytelling
language: - en
Ministral-3B-Heresy-GGUF (Unfiltered Edition)
This repository contains GGUF quantizations of the Ministral-3B-Heresy model. This model is a specialized fine-tune designed to bridge the gap between small-parameter efficiency and high-intensity, unfiltered narrative roleplay.
⚠️ CONTENT WARNING: NO FILTERS
This model has been specifically conditioned on "Heresy" datasets. It is designed to ignore standard safety guardrails and "assistant-like" refusals. It will generate visceral, intense, and explicit content without hesitation.
Model Highlights
- Zero Refusals: Optimized to bypass moralizing "As an AI..." interjections.
- Visceral Narrative: Focused on physical sensations, sensory details, and raw emotional dialogue.
- Efficient Size: At only 3B parameters, this model runs at high speed on consumer GPUs and mobile devices while maintaining 7B-level roleplay logic.
Quantization Details
All files were quantized using llama.cpp to ensure compatibility with the latest GGUF standards.
| File | Size | Recommended Use Case |
|---|---|---|
ministral-fp16.gguf |
~6.5 GB | For further quantization or maximum fidelity. |
ministral-Q8_0.gguf |
~3.5 GB | Maximum quality with no measurable loss. |
ministral-Q6_K.gguf |
~2.7 GB | High fidelity; nearly indistinguishable from FP16. |
ministral-Q5_K_M.gguf |
~2.3 GB | Best Balance for speed and logic. |
ministral-Q4_K_M.gguf |
~2.1 GB | Lightweight; ideal for mobile or 4GB VRAM. |
Recommended Inference Settings
To prevent the model from defaulting to "polite" base-model behavior, use the following parameters:
- Sampling: - Temperature:
0.85- Repeat Penalty:
1.18(Critical to prevent loops in descriptive scenes) - Top P:
0.9 - Top K:
40
- Repeat Penalty:
- System Prompting: > Always use a system prompt to define your character. The model excels when told to be "descriptive," "visceral," and "unfiltered."
Technical Background
This quantization project was executed in a CUDA-accelerated environment. Special attention was paid to preserving the custom tokenization required for the Ministral architecture, ensuring that multi-turn conversations remain stable and coherent.
Disclaimer
The creators of this quantization are not responsible for the content generated by the model. By downloading this model, you acknowledge the unfiltered nature of the underlying weights.