georgehenney/Qwen3-VL-4B-Instruct-heretic-8refusal

Overview

georgehenney/Qwen3-VL-4B-Instruct-heretic-8refusal is a 4 billion parameter vision-language model based on the Qwen3-VL series, specifically a 'decensored' version of Qwen/Qwen3-VL-4B-Instruct. This model has been fine-tuned using the Heretic v1.0.1 tool to significantly reduce its refusal rate from 92/100 to 8/100, offering a less restrictive response generation profile compared to its original counterpart. It maintains the core multimodal capabilities of the Qwen3-VL architecture, which is designed for advanced visual perception and reasoning.

Key Capabilities

• Reduced Refusal Rate: Engineered to provide responses in scenarios where the original Qwen3-VL-4B-Instruct might refuse, with a refusal rate of only 8% on tested prompts.
• Comprehensive Multimodal Understanding: Inherits the Qwen3-VL's ability to process and understand both visual and textual inputs, including advanced spatial perception, long context video understanding, and enhanced multimodal reasoning.
• Visual Agent Functionality: Capable of operating PC/mobile GUIs, recognizing elements, understanding functions, and completing tasks.
• Advanced OCR: Supports 32 languages and performs robustly in challenging conditions (low light, blur, tilt), with improved parsing of long document structures.
• Text Understanding: Offers seamless text-vision fusion for unified comprehension, with text understanding on par with pure LLMs.

Good for

• Applications requiring a less restrictive or 'decensored' vision-language model for broader content generation.
• Tasks involving complex visual and textual reasoning, such as image description, visual question answering, and document analysis, where the original model's safety filters might be too stringent.
• Developing visual agents that interact with graphical user interfaces.
• Use cases demanding robust OCR across multiple languages and challenging image conditions.