DreamFast/qwen3-4b-heretic

Hugging Face
TEXT GENERATIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Mar 10, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

DreamFast/qwen3-4b-heretic is a 4 billion parameter language model, an 'abliterated' version of Qwen/Qwen3-4B created using Heretic v1.2.0. This model significantly reduces refusals while maintaining original model quality, making it highly suitable as an uncensored text encoder for image generation models like Z-Image and FLUX.2 Klein 4B. It is available in various ComfyUI-native quantized formats, all produced with SVD-guided learned rounding for maximum fidelity.

Loading preview...

Overview

DreamFast/qwen3-4b-heretic is an 'abliterated' version of the Qwen 3 4B base model, specifically engineered to reduce refusal behaviors. Created using the Heretic v1.2.0 tool, this model underwent 200 optimization trials, resulting in a version with only 3 refusals out of 100, a significant reduction from the original model's 100/100 refusals. Crucially, this refusal reduction was achieved with zero measurable KL divergence, indicating no damage to the model's core capabilities.

Key Features & Optimizations

  • Reduced Refusals: Achieves 3/100 refusals compared to 100/100 in the base model, without compromising quality.
  • Zero Model Damage: KL Divergence of 0.0000 confirms the abliteration process surgically removed refusal mechanisms without impacting other functionalities.
  • Extensive Quantization: Provided in multiple ComfyUI-native quantized formats (FP8, INT8, INT4, NVFP4, MXFP8) and GGUF formats (Q8_0, Q6_K, Q5_K_M, Q4_K_M, Q3_K_M).
  • High-Fidelity Quantization: All quantized variants utilize SVD-guided learned rounding (AdaRound via convert_to_quant) to minimize output reconstruction error, offering higher fidelity than naive rounding.

Ideal Use Cases

  • Uncensored Text Encoding for Image Generation: Specifically designed to serve as an uncensored text encoder for models like Z-Image and FLUX.2 Klein 4B.
  • ComfyUI Workflows: Natively loads in ComfyUI 0.30.0+ without plugins, supporting various hardware configurations with optimized quantized formats.
  • Applications Requiring Reduced Refusal: Suitable for scenarios where a more permissive language model is desired without sacrificing core performance.

Popular Sampler Settings

Top 3 parameter combinations used by Featherless users for this model. Click a tab to see each config.

temperature
top_p
top_k
frequency_penalty
presence_penalty
repetition_penalty
min_p