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

DreamFast/gemma-3-12b-it-heretic-v2 Overview

This model is an abliterated version of Google's Gemma 3 12B IT, created using Heretic v1.3.0. Its primary distinction is a significant reduction in content refusals (from 100/100 to 8/100 in trials) while preserving model quality, indicated by a low KL divergence of 0.0801. This makes it particularly effective for applications requiring less censorship, such as an uncensored text encoder for video generation models like LTX-2.

Key Capabilities & Features

• Reduced Refusals: Achieves 92% reduction in refusals compared to the base model, enabling more faithful prompt encoding for creative content.
• Vision Preserved: All ComfyUI variants retain vision_model and multi_modal_projector keys, supporting I2V (image-to-video) prompt enhancement.
• Diverse Quantization: Offered in five formats: FP8, INT8 (ConvRot), NVFP4, MXFP8, and GGUF, catering to a wide range of GPU hardware from Ada to Blackwell.
• INT8 (ConvRot): Near-lossless INT8 quantization via SVD-guided learned rounding, compatible with any modern GPU (Ampere+) and natively supported in ComfyUI v0.27.0+.
• MXFP8 & NVFP4: Specialized quantization for Blackwell GPUs, offering efficient performance and smaller file sizes.

Ideal Use Cases

• Video Generation (LTX-2): Optimized as a text encoder for LTX-2, where it helps in generating videos with more faithful and less censored prompt adherence.
• Creative Content Generation: Suitable for applications where reduced censorship in text embeddings is desired, leading to less sanitized or altered outputs.
• Resource-Constrained Environments: The variety of quantization formats allows deployment across different hardware, from high-end Blackwell GPUs to older Ampere+ cards, optimizing for VRAM and inference speed.