CCSSNE/DreamFast-qwen3-4b-heretic

Hugging Face
TEXT GENERATIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 23, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

CCSSNE/DreamFast-qwen3-4b-heretic is a 4 billion parameter Qwen 3 base model, developed by CCSSNE/DreamFast, that has been 'abliterated' using Heretic v1.2.0 to significantly reduce refusals while maintaining model quality. This model is specifically optimized as an uncensored text encoder for image generation models like Z-Image and FLUX.2 Klein 4B, offering a 32768 token context length. It achieves a refusal rate of 3/100 compared to the original's 100/100, with zero measurable KL divergence, indicating no damage to its core capabilities.

Loading preview...

Overview

CCSSNE/DreamFast-qwen3-4b-heretic is a 4 billion parameter language model based on Qwen 3, specifically modified to reduce content refusals. This "abliterated" version was created using the Heretic v1.2.0 tool, which surgically removed the refusal mechanism without measurable damage to the model's core capabilities (KL divergence of 0.0000).

Key Differentiators

  • Reduced Refusals: Achieves a refusal rate of 3/100, a significant reduction from the base Qwen 3 4B's 100/100, making it less prone to censorship.
  • Quality Preservation: The abliteration process resulted in zero measurable KL divergence, ensuring that the model's original quality and capabilities are maintained.
  • Optimized for Image Generation: Primarily designed as an uncensored text encoder, making it highly suitable for integration with image generation models such as Z-Image and FLUX.2 Klein 4B.

Usage and Formats

The model is available in various formats to support different use cases:

  • HuggingFace Format: Compatible with the transformers library for general NLP tasks.
  • ComfyUI Format: Optimized for direct use within ComfyUI workflows, with options for bf16, FP8, and NVFP4 quantizations.
  • GGUF Format: Supports llama.cpp and ComfyUI-GGUF, offering multiple quantization levels (e.g., Q4_K_M recommended for balance).

Limitations

This model inherits all limitations of the base Qwen 3 4B model. While refusals are significantly reduced, they are not entirely eliminated (3/100 refusals remain).

Finetunes
1 models