This repository hosts GGUF weights and the associated vision projection matrix for Ministral-3-8B-Reasoning-2512-heretic, quantized from the source floating-point tensors provided by coder3101/Ministral-3-8B-Reasoning-2512-heretic.

🔄 Sister Repository: Check out the Imatrix Sister Repository for enhanced precision at lower bit fractions.

[!NOTE] If you plan on using 4-bit or 5-bit variants, consider the imatrix sister repository instead — importance matrix calibration improves logic retention at those bit depths. This repository is best suited if you want the near-lossless Q8_0 build.

ℹ️ Model Profile & Core Features

Ministral-3-8B-Reasoning-2512 is a vision-language model by Mistral AI from the Ministral 3 family, built for edge deployment. It combines an 8.4B language model with a 0.4B vision encoder for multimodal understanding, and is the reasoning post-trained variant — specifically optimised for math, coding, STEM, and complex multi-step reasoning. The model supports multilingual input, native function calling, JSON output, and best-in-class agentic capabilities at its scale.

The heretic suffix denotes post-processing via the Heretic v1.1.0 abliteration framework performed by coder3101, using the direction_index method to suppress refusal vectors while preserving reasoning chains and multimodal capabilities.

📋 Technical Specifications

Property	Value
Base Architecture	Ministral-3 (8.4B LM + 0.4B Vision Encoder)
Developed by	Mistral AI
Primary Use	Reasoning, math, STEM, vision, agentic tasks
Context Window	262,144 tokens
Vision Encoder	0.4B (integrated, early-fusion)
Languages	English, French, Spanish, German, Italian, Portuguese, Dutch, Chinese, Japanese, Korean, Arabic
Abliteration Tool	Heretic v1.1.0
Abliteration Method	direction_index (single-direction refusal suppression)
Prompt Format	ChatML

🛠️ Heretic Overrides (ARA)

Property	Value
direction_index	15.03
attn.o_proj.max_weight	1.17
attn.o_proj.max_weight_position	20.88
attn.o_proj.min_weight	0.55
attn.o_proj.min_weight_distance	2.72
mlp.down_proj.max_weight	1.49
mlp.down_proj.max_weight_position	23.45
mlp.down_proj.min_weight	1.40
mlp.down_proj.min_weight_distance	18.55

📊 Refusal Bypass Metrics

[!NOTE] The metrics below are self-reported by the original model author (coder3101) and have not been independently reproduced.

Metric	This model	Original (mistralai/Ministral-3-8B-Reasoning-2512)
KL divergence	0.0712	0 (by definition)
Refusals	4/100	96/100

🧮 Numerical & Tensor Formats

Property	Value
Text Tensors	Q4_K_M, Q5_K_M, Q8_0
Vision Tensors	Q8_0, BF16

📦 Available Model Files

Main model weights

Filename	Quantization	llama.cpp Build	Size	Download
`Ministral-3-8B-Reasoning-2512-heretic-Q4_K_M.gguf`	`Q4_K_M`	`b9803`	4.84 GB	📥 Download
`Ministral-3-8B-Reasoning-2512-heretic-Q5_K_M.gguf`	`Q5_K_M`	`b9870`	5.64 GB	📥 Download
`Ministral-3-8B-Reasoning-2512-heretic-Q8_0.gguf`	`Q8_0`	`b9870`	8.41 GB	📥 Download

mmproj — vision projector files

Filename	Quantization	Size	Download
`mmproj-Ministral-3-8B-Reasoning-2512-heretic-Q8_0.gguf`	`Q8_0`	448 MB	📥 Download
`mmproj-Ministral-3-8B-Reasoning-2512-heretic-BF16.gguf`	`BF16`	827 MB	📥 Download

🎛️ Component Pairing Guide

Download exactly one main weights file:

Q4_K_M: Balanced 4-bit format suitable for most everyday use.
Q5_K_M: Higher-fidelity mid-range format recommended as a general default.
Q8_0: Near-lossless 8-bit format for when memory is not a constraint.

mmproj files (optional): multimodal vision projectors. Pass one via the --mmproj flag in llama.cpp to enable image input.

BF16 (Recommended): Highest possible image processing accuracy. While older projectors were small, modern vision towers can hover around 1GB. If you are tight on VRAM, it is completely viable to run this on system RAM (CPU) with a minimal performance penalty, saving your precious GPU space for the main model layers.
Q8_0: Cuts the projector file size and memory footprint in half (~500MB for larger 1GB files). Use this if you prefer to keep the vision tower hosted entirely on your GPU but need to claw back some VRAM to avoid Out-Of-Memory (OOM) crashes.

⚡ Deployment & Execution Commands

[!IMPORTANT] The vision projector (--mmproj) must be supplied at runtime whenever image inputs are used. Omitting it disables multimodal capability entirely.

[!NOTE] Mistral AI recommends the following sampling configuration for best results: temperature=0.7, top_p=0.95.

[!TIP] Swap the -m filename below for either quantized file depending on your size/quality trade-off preference.

llama.cpp CLI (with image)

bash

./llama-cli \
  -m Ministral-3-8B-Reasoning-2512-heretic-Q4_K_M.gguf \
  --mmproj mmproj-Ministral-3-8B-Reasoning-2512-heretic-Q8_0.gguf \
  -c 8192 \
  -ngl 99 \
  --image "path/to/image.jpg" \
  -p "<|im_start|>user\nSolve the problem shown in the image step by step.<|im_end|>\n<|im_start|>assistant\n"

OpenAI-Compatible API Server

bash

./llama-server \
  --host 0.0.0.0 \
  --port 8080 \
  -m Ministral-3-8B-Reasoning-2512-heretic-Q4_K_M.gguf \
  --mmproj mmproj-Ministral-3-8B-Reasoning-2512-heretic-Q8_0.gguf \
  -c 16384 \
  -ngl 99 \
  --flash-attn

💬 Chat Templates & Prompt Design (ChatML)

text

<|im_start|>system
You are an expert reasoning assistant. Think step by step.<|im_end|>
<|im_start|>user
Your question or image payload here.<|im_end|>
<|im_start|>assistant

⚠️ Safety & Operational Notes

This model is abliterated and will generate content that standard aligned models refuse. Use responsibly and in compliance with applicable laws.
Designed for edge deployment; fits in 24 GB VRAM at BF16 and under 12 GB when quantized.
Vision tensors are kept at BF16 or Q8_0 depending on the mmproj variant chosen, to prevent degradation of spatial and symbolic features in diagrams and formulas.
For better output quality at this quantization level, consider the imatrix variant in the companion repository.

FadedRedStar/Ministral-3-8B-Reasoning-2512-heretic-GGUF

🤖 Ministral-3-8B-Reasoning-2512-heretic — GGUF