coder3101/Qwen3-VL-4B-Thinking-heretic

VISIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jan 17, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The coder3101/Qwen3-VL-4B-Thinking-heretic is a 4 billion parameter vision-language model, based on the Qwen3-VL-4B-Thinking architecture developed by Qwen, with a 32768 token context length. This version is a decensored variant, created using Heretic v1.1.0, demonstrating significantly reduced refusal rates compared to the original model. It excels in multimodal reasoning, visual agent capabilities, and advanced spatial perception, making it suitable for tasks requiring robust visual and textual understanding without content restrictions.

Loading preview...

Model Overview

The coder3101/Qwen3-VL-4B-Thinking-heretic is a 4 billion parameter vision-language model, derived from the Qwen3-VL-4B-Thinking architecture by Qwen. This specific iteration has been decensored using Heretic v1.1.0, resulting in a notable reduction in refusal rates (3/100 compared to the original's 97/100), while maintaining a low KL divergence of 0.0259 from the base model. It features a substantial 32768 token context length, enabling comprehensive understanding of long-form text and complex visual inputs.

Key Capabilities

  • Decensored Performance: Significantly reduced content refusals, offering more direct and unfiltered responses.
  • Visual Agent: Capable of operating PC/mobile GUIs, recognizing elements, understanding functions, and completing tasks.
  • Advanced Spatial Perception: Judges object positions, viewpoints, and occlusions, providing stronger 2D and enabling 3D grounding for spatial reasoning.
  • Enhanced Multimodal Reasoning: Excels in STEM/Math tasks, performing causal analysis and delivering logical, evidence-based answers.
  • Upgraded Visual Recognition: Broad and high-quality pretraining allows recognition of a wide array of entities including celebrities, anime, products, and landmarks.
  • Expanded OCR: Supports 32 languages and is robust in challenging conditions (low light, blur, tilt), with improved parsing for rare characters and long document structures.
  • Text Understanding: Achieves text understanding on par with pure LLMs, ensuring seamless text-vision fusion.

Good For

  • Applications requiring a vision-language model with fewer content restrictions.
  • Tasks involving visual agents, GUI interaction, and tool invocation.
  • Complex multimodal reasoning, especially in STEM and mathematical domains.
  • Scenarios demanding advanced spatial understanding and 3D grounding.
  • OCR tasks across multiple languages and challenging image conditions.

Finetunes
1 models