fsxedx/Qwen3-4b-Z-Image-Turbo-AbliteratedV1

TEXT GENERATIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 24, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The Qwen3-4b-Z-Image-Turbo-AbliteratedV1 is a 4 billion parameter text encoder developed by BennyDaBall, specifically engineered to reduce refusal rates in image generation and general conversational contexts. This model utilizes an "abliterated" p-e-w heretic method, refined over 1000 trials, resulting in a minimal KL Divergence of 0.0004 and a refusal rate of only 4/100 in torture tests. It is designed for developers seeking a highly permissive model for creative and unconstrained text-to-image prompts and general text generation, offering a 32768 token context length.

Loading preview...

Overview

Qwen3-4b-Z-Image-Turbo-AbliteratedV1, developed by BennyDaBall, is a 4 billion parameter text encoder designed to significantly reduce refusal rates in both image generation prompts and general conversational scenarios. This model was created using an "abliterated" p-e-w heretic method, refined through 1000 trials to specifically target and minimize refusals.

Key Capabilities

  • Extremely Low Refusal Rate: Achieves a refusal rate of only 4/100 in rigorous testing, making it highly permissive for diverse content generation.
  • Minimal "Lobotomy" Effect: Demonstrates a very low KL Divergence of 0.0004, indicating that its core capabilities remain largely intact despite the refusal-reduction modifications.
  • Optimized for Image Generation: Specifically targets and overcomes common refusals encountered when generating image prompts.
  • General Purpose Permissiveness: Also addresses and reduces general conversational refusals, allowing for broader application.

Good For

  • Developers and creators who require a highly unconstrained model for text-to-image applications.
  • Use cases where traditional LLMs might refuse to generate certain content due to safety filters.
  • Experimentation with creative and boundary-pushing text generation without frequent content restrictions.

Available Formats

The model is provided in various GGUF quantization formats, ranging from F16 (8.05 GB) for full precision to Q2_K (1.67 GB) for minimal size, offering flexibility for different hardware and performance needs.

Finetunes
2 models