👹 LittleMonster — 12B Reasoner (v2)

A dynamically reasoning, restriction-free 12B language model built on Gemma 3

[!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

💡 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

Training Data

🗣️ Feedback

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