Overview

DreamFast/qwen3-8b-heretic is an 8 billion parameter language model derived from Qwen's Qwen3-8B, specifically processed using the Heretic v1.2.0 abliteration tool. The primary goal of this modification is to significantly reduce model refusals from 100/100 to 13/100, making it more permissive while preserving the original model's quality, as indicated by a low KL Divergence of 0.0838.

Key Capabilities & Features

• Reduced Refusals: Achieves an 87% reduction in refusal rates compared to the base Qwen3-8B model.
• High Quality: Abliteration process maintains model integrity with minimal damage.
• Optimized for Image Generation: Designed to function effectively as an uncensored text encoder, particularly for models like Klein 9B.
• Flexible Formats: Provided in HuggingFace, ComfyUI (bf16, FP8, NVFP4), and GGUF (various quantizations including Q4_K_M recommended) formats.
• NVFP4 Quantization: Offers highly efficient NVFP4 (4-bit floating point) variants, ideal for Blackwell GPUs (RTX 5090/5080) with native FP4 tensor cores, and supported on older GPUs via software dequantization.

Recommended Use Cases

• Uncensored Text Encoding: Ideal for applications requiring a less restrictive text encoder, especially in conjunction with image generation models.
• Creative Content Generation: Suitable for scenarios where the base model's refusal behavior might hinder creative or open-ended text generation.
• Resource-Efficient Deployment: GGUF and NVFP4 formats enable deployment on systems with varying hardware constraints, from high-end GPUs to low VRAM setups.