LVMCS/Gemma-3-12B-IT-Heretic-v2-Abliterated-Comfy

Hugging Face
VISIONConcurrency Cost:1Model Size:12BQuant:FP8Ctx Length:32kPublished:Mar 13, 2026License:gemmaArchitecture:Transformer Warm

LVMCS/Gemma-3-12B-IT-Heretic-v2-Abliterated-Comfy is a 12 billion parameter instruction-tuned Gemma 3 model, created by LVMCS, that has been 'abliterated' using Heretic v1.2.0. This process significantly reduces model refusals while preserving quality, making it suitable as an uncensored text encoder for video generation models like LTX-2. It maintains vision capabilities and offers various quantization formats, including NVFP4 for ComfyUI and GGUF for llama.cpp, optimized for diverse hardware and workflows.

Loading preview...

LVMCS/Gemma-3-12B-IT-Heretic-v2-Abliterated-Comfy Overview

This model is an 'abliterated' version of Google's Gemma 3 12B IT, processed using Heretic v1.2.0. The abliteration significantly reduces refusals (from 100/100 to 8/100) while maintaining model quality, as indicated by a low KL divergence of 0.0801. It is specifically designed to function as an uncensored text encoder, particularly for video generation models like LTX-2, by removing soft censorship in embeddings.

Key Capabilities & Features

  • Reduced Refusals: Achieves 92% reduction in refusals compared to the base model, allowing for more faithful prompt encoding.
  • Vision Preserved: Includes vision_model and multi_modal_projector keys, supporting I2V (image-to-video) prompt enhancement in ComfyUI.
  • Optimized Formats: Available in HuggingFace, ComfyUI (bf16, FP8, NVFP4), and GGUF (various quantizations like Q4_K_M) formats for broad compatibility.
  • ComfyUI Integration: Supports native ComfyUI quantization (NVFP4) for Blackwell GPUs and older GPUs via software dequantization.

Use Cases & Considerations

  • LTX-2 Text Encoding: Ideal for use as a text encoder in LTX-2 workflows, providing less censored embeddings for creative content generation.
  • Uncensored Text Generation: While abliteration removes refusals, the model's knowledge is limited to its original training data; it won't generate content it was never exposed to.
  • Multi-GPU/CPU Offloading: Compatible with optimized LTX-2 workflows for multi-GPU setups or CPU offloading.
  • Limitations: Inherits base Gemma 3 12B limitations; abliteration reduces, but does not eliminate, all refusals.

Finetunes
1 models