Leon1000/qwen3-4b

TEXT GENERATIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 4, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Leon1000/qwen3-4b is a 4 billion parameter Qwen 3 base model that has been 'abliterated' using Heretic v1.2.0 to significantly reduce refusals while maintaining original model quality. This model is specifically optimized as an uncensored text encoder for image generation models like Z-Image and FLUX.2 Klein 4B. It achieves a refusal rate of 3/100 compared to the original 100/100, with zero measurable KL divergence, indicating no damage to its core capabilities.

Loading preview...

Overview

Leon1000/qwen3-4b is an 'abliterated' version of the Qwen 3 4B base model, processed with Heretic v1.2.0. The primary goal of this modification is to drastically reduce the model's refusal rate without compromising its original quality or capabilities. Through 200 optimization trials, a specific trial (Trial 96) was selected, resulting in a refusal rate of just 3/100, a significant improvement over the base model's 100/100 refusals. Crucially, this abliteration process resulted in zero measurable KL divergence, confirming that the refusal mechanism was surgically removed with no damage to the model's core functions.

Key Capabilities

  • Reduced Refusals: Achieves a refusal rate of 3/100, making it highly suitable for applications requiring less restrictive content generation.
  • Maintained Quality: The abliteration process ensures no measurable damage to the model's original capabilities, preserving its performance.
  • 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.
  • Versatile Formats: Available in HuggingFace, ComfyUI (bf16, FP8, NVFP4), and GGUF (various quantizations including Q4_K_M recommended) formats for broad compatibility.

Good For

  • Developers and artists using ComfyUI with image generation models like Z-Image or FLUX.2 Klein 4B who need a less restrictive text encoder.
  • Applications requiring a 4B parameter language model with significantly reduced content refusals.
  • Experimentation with abliterated models and understanding the impact of refusal reduction techniques.

Finetunes
1 models