zhangsanfeng364/gemma-3-12b-it-heretic-v2

Model Overview

This model, gemma-3-12b-it-heretic-v2, is an abliterated version of Google's 12 billion parameter Gemma 3 instruction-tuned model. It was created using the Heretic v1.2.0 tool, specifically designed to reduce model refusals while preserving overall quality. The abliteration process involved 200 trials, with Trial 174 selected for achieving 8/100 refusals (compared to 100/100 for the original) at a low KL divergence of 0.0801, indicating minimal damage to the base model.

Key Differentiators & Capabilities

• Reduced Refusals: Significantly lowers the model's tendency to refuse prompts, making it more permissive for creative content generation.
• Vision Preservation: Unlike some abliterated models, this version retains vision_model and multi_modal_projector keys, enabling its use for Image-to-Video (I2V) prompt enhancement in frameworks like ComfyUI with LTX-2.
• Optimized for Video Generation: Primarily intended as an uncensored text encoder for video generation models, addressing the issue of "soft censorship" in embeddings that can weaken adherence to creative prompts.
• Quantization Options: Available in various formats including HuggingFace safetensors, ComfyUI-native bf16, fp8, and nvfp4 (optimized for Blackwell GPUs), and multiple GGUF quantizations for llama.cpp and ComfyUI-GGUF.

Usage & Limitations

This model can be integrated with transformers, ComfyUI (especially for LTX-2 workflows), ComfyUI-GGUF, and llama.cpp. While abliteration removes soft censorship, the model's knowledge is still limited to its original training data. It does not gain knowledge of content it was never exposed to, and its impact on LTX-2 is nuanced, as LTX-2 was trained on original Gemma embeddings. NVFP4 quantization performs best on Blackwell GPUs, though software dequantization supports older hardware.