slowface/Qwen3-VL-32B-Gemini-Heretic-Uncensored-Thinking

Hugging Face
VISIONConcurrent Unit Cost:2Model Size:33.4BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 22, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Warm

The slowface/Qwen3-VL-32B-Gemini-Heretic-Uncensored-Thinking is a 33.4 billion parameter vision-language model, based on the Qwen3-VL architecture, specifically fine-tuned for uncensored content generation and enhanced Gemini-style reasoning. It features a 32768 token context length and is designed to accept all image types, providing detailed analytics and output generation. This model prioritizes compact, precise reasoning and broad content generation without refusal, making it suitable for diverse and unrestricted applications.

Loading preview...

Model Overview

The slowface/Qwen3-VL-32B-Gemini-Heretic-Uncensored-Thinking is a 33.4 billion parameter vision-language model derived from the Qwen3-VL-32B-Thinking base, developed by slowface. This iteration focuses on providing completely uncensored content generation and integrating a "Gemini thinking" style for highly compact and precise reasoning. It maintains the core functionalities and metrics of the original Qwen architecture while significantly reducing content refusals.

Key Capabilities

  • Uncensored Content Generation: Designed to generate any kind of content without restrictions or refusals, accepting all image types.
  • Enhanced Reasoning: Features a "Gemini thinking" style, characterized by very compact (4-6 paragraphs or less) yet detailed and precise reasoning, particularly for image analytics.
  • Vision-Language Integration: Leverages the Qwen3-VL base's advanced visual perception, reasoning, and multimodal capabilities, including visual agent operations, visual coding boost, and advanced spatial perception.
  • Extended Context: Supports a native 256K context length (though the model card states 32768 tokens), enabling processing of long documents and video content.
  • Robust OCR: Offers expanded OCR capabilities supporting 32 languages, with improved performance in challenging conditions and for rare characters.
  • Low Refusal Rate: Demonstrates a significantly reduced refusal rate (3/100) compared to the original model (97/100), with a low KL divergence of 0.0048 indicating minimal damage to the base model during de-censoring.

Ideal Use Cases

This model is particularly suited for applications requiring:

  • Unrestricted Content Creation: For users needing to generate diverse content without built-in censorship.
  • Detailed Image Analysis: Where precise and compact reasoning about visual inputs is critical.
  • Broad Multimodal Tasks: Leveraging its comprehensive visual and text understanding for various applications, including visual agents and coding from images.
  • Research and Development: For exploring the boundaries of AI capabilities without content limitations.

Finetunes
1 models