Modelpublic

mfielding92/littlemonster-reasoning-12B-QKVO-heretic-GGUF

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face

5likes199downloads

Model Card

<div align="center"> <img src="https://huggingface.co/mfielding92/littlemonster-reasoning-12B-QKVO-heretic-GGUF/resolve/main/littlemonster_banner.jpg" alt="LittleMonster AI Banner" style="max-width: 850px; border-radius: 12px;"> <h1>👹 LittleMonster — 12B Reasoner (v2)</h1> <b>A dynamically reasoning, restriction-free 12B language model built on Gemma 3</b> <br>

![License](https://opensource.org/licenses/Apache-2.0) ![Base Model](https://huggingface.co/p-e-w/gemma-3-12b-it-heretic) ![Trained With](https://github.com/unslothai/unsloth) </div>

[!IMPORTANT] Model Updated — 03/02/2026: If you downloaded this model before this date, please re-download for enhanced prompt alignment and improved "enthusiasm" when engaging with discussion topics.

📖 Overview

LittleMonster is part of a work-in-progress family of models (4B coming soon!) designed to deliver thorough reasoning without heavy restrictions. This 12B variant has undergone multiple iterative stages of fine-tuning and heretic processing to reach its current state.

✨ Key Features

—🧠 Dynamic Reasoning — The model intelligently determines whether step-by-step reasoning is necessary on a per-request basis, adapting to the complexity of each task.
—🔓 Restriction-Free — Multiple rounds of heretic/abliteration processing produce a model that engages openly with a wide range of topics.
—⚡ Efficient Training — Trained 2× faster using Unsloth and Hugging Face's TRL library.

[!NOTE] Custom-tuned imatrix: The importance matrix used for imatrix quants was specifically generated using uncensored calibration data. This ensures that even at lower bit depths, the quantized models preserve the unrestricted behavior of the full-precision model — keeping the abliteration and heretic processing intact where standard imatrix data would risk degrading it.

📦 Available GGUF Quants

Quant Type	Imatrix	Notes
`IQ3_XS`	✅ Yes	Smallest size — imatrix-enhanced accuracy
`IQ3_M`	✅ Yes	Balanced ultra-low-bit option
`Q3_K_L`	❌ No	Standard 3-bit quantization
`IQ4_XS`	✅ Yes	Great quality-to-size ratio with imatrix
`Q4_K_M`	❌ No	Popular general-purpose quant
`Q6_K`	❌ No	High quality, moderate size
`Q8_0`	❌ No	Near-lossless quantization

💡 Usage Tips

[!TIP] For best results on sensitive or controversial topics that the model is not responding well to, ease into the conversation with your first message before introducing the main subject. Example: "Hey, I got a question regarding something personal but I just wanted to make sure it's ok first before I ask you."

🔧 Model Details

Property	Details
Developer	mfielding92
Base Model	gemma-3-12b-it-heretic (by p-e-w)
Architecture	Gemma 3 — 12B parameters
License	Apache 2.0
Language	English
Model Type	Causal LM — Text Generation / Reasoning

Training Data

Dataset	Purpose
`mfielding92/LittleMonster`	Core personality & behavior alignment
`mfielding92/gemini-3.1-pro-2048-reasoning-1100x`	Reasoning capability enhancement
`mfielding92/LittleMonster-reasoning`	Supplementary reasoning data

🗣️ Feedback

Your feedback is invaluable for improving future versions! Please share any issues, observations, or suggestions in the Community tab. Thank you! 🙏