Gemma 3 12B IT - Heretic v2 (Abliterated)

This model is an "abliterated" version of Google's Gemma 3 12B IT, created using Heretic v1.2.0. Its primary distinction is a significant reduction in refusals, making it an uncensored text encoder while maintaining minimal model damage (KL Divergence of 0.0801). This version is particularly suited for use in video generation models like LTX-2, where it helps ensure more faithful prompt encoding for creative content by removing soft censorship in embeddings.

Key Capabilities

• Reduced Refusals: Achieves 8/100 refusals on previously censored prompts, down from 100/100 in the original model.
• Preserved Quality: Maintains model quality with a low KL Divergence of 0.0801.
• Vision Support: ComfyUI variants retain vision_model and multi_modal_projector keys, enabling I2V (image-to-video) prompt enhancement.
• Optimized Quantization: Includes NVFP4 quantization for Blackwell GPUs, offering ~3x smaller file sizes than bf16, and updated GGUF support.

Good for

• Uncensored Text Encoding: Ideal for applications requiring a text encoder with reduced content restrictions.
• Video Generation (LTX-2): Specifically designed to enhance prompt adherence and creative output in LTX-2 text-to-video (T2V) and image-to-video (I2V) workflows.
• Resource-Efficient Deployment: Offers various quantized formats (NVFP4, FP8, GGUF Q4_K_M) for deployment on diverse hardware, including Blackwell GPUs and systems with limited VRAM.