DreamFast/qwen3-4b-heretic

Overview

DreamFast/qwen3-4b-heretic is a specialized version of the Qwen 3 4B language model, processed using the Heretic v1.2.0 tool. The primary goal of this 'abliteration' was to drastically reduce the model's refusal rates without compromising its core quality or capabilities. Through 200 optimization trials, a specific configuration (Trial 96) was selected, achieving a reduction in refusals from 100/100 to just 3/100, with zero measurable KL divergence, ensuring no model damage.

Key Capabilities & Features

• Reduced Refusals: Significantly lowers the model's tendency to refuse prompts, making it more versatile for various applications.
• Quality Preservation: Maintains the original Qwen 3 4B model's quality, as indicated by zero KL divergence.
• Optimized for Image Generation: Specifically designed to function as an uncensored text encoder for advanced image generation models such as Z-Image and FLUX.2 Klein 4B.
• Multiple Formats: Available in HuggingFace, ComfyUI, and GGUF formats, with various quantization options (bf16, fp8, nvfp4, Q8_0, Q6_K, Q5_K_M, Q4_K_M, Q3_K_M) for flexibility across different hardware and use cases.
• NVFP4 Support: Includes NVFP4 variants optimized for ComfyUI, offering significant size reduction and native loading, with potential performance benefits on Blackwell GPUs.

When to Use This Model

• Uncensored Text Encoding: Ideal for users requiring a text encoder with minimal refusals for image generation tasks.
• Image Generation Workflows: Directly compatible with ComfyUI for integration into Z-Image or FLUX.2 Klein 4B workflows.
• Resource-Constrained Environments: The availability of highly quantized GGUF and ComfyUI formats makes it suitable for deployment on devices with limited VRAM.
• Experimentation with Abliteration: Useful for developers interested in the practical application of tools like Heretic for model modification.