DavidAU/Qwen3-VL-8B-GLM-4.7-Flash-Heretic-Uncensored-Thinking

VISIONConcurrency Cost:1Model Size:8BQuant:FP8Ctx Length:32kPublished:Jan 30, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

DavidAU/Qwen3-VL-8B-GLM-4.7-Flash-Heretic-Uncensored-Thinking is an 8 billion parameter vision-language model, based on Qwen3-VL-8B-Instruct, enhanced with GLM 4.7 Flash thinking/reasoning capabilities. It features complete uncensoring and detailed reasoning traces, preserving Qwen's core functions and metrics. This model is designed for all use cases, excelling in image analytics and output generation without content restrictions, and supports a 256K context length.

Loading preview...

Model Overview

DavidAU/Qwen3-VL-8B-GLM-4.7-Flash-Heretic-Uncensored-Thinking is an 8 billion parameter vision-language model derived from Qwen3-VL-8B-Instruct. Its primary distinction lies in its complete uncensoring and the integration of GLM 4.7 Flash thinking/reasoning capabilities, which are implanted with minimal power to preserve Qwen's original functions and metrics. This model aims to provide detailed and compact reasoning, particularly enhancing image analytics and output generation.

Key Capabilities & Enhancements

  • Uncensored Content Generation: Designed to generate any kind of content and accept all image types without restrictions.
  • Enhanced Reasoning: Integrates GLM 4.7 Flash thinking, significantly improving image analytics, detail extraction, and output generation. This reasoning is noted as "TEMP stable."
  • Performance Improvement: Benchmarks show notable improvements over the non-thinking Qwen3-VL-8B-Instruct-heretic-qx86-hi, particularly in arc_challenge (0.572 vs 0.437) and hellaswag (0.716 vs 0.526).
  • De-censoring Metrics: Achieves a KL divergence of 0.1738 (indicating minimal damage to the base model) and significantly reduced refusals (6/100 compared to 100/100 for the original).
  • Base Model Features: Inherits advanced multimodal capabilities from the Qwen3-VL-8B-Instruct base, including visual agent functionality, enhanced spatial perception, long context (256K native, expandable to 1M), improved multimodal reasoning for STEM/Math, upgraded visual recognition, and expanded OCR supporting 32 languages.

Good For

  • Applications requiring unrestricted content generation: Ideal for use cases where content filtering or censorship is undesirable.
  • Detailed image analysis and reasoning: Benefits from the integrated GLM 4.7 Flash thinking for in-depth visual perception.
  • Developers seeking a highly capable, uncensored VL model: Offers a unique combination of advanced vision-language understanding with complete freedom in output.
  • Research into uncensored model behavior and reasoning integration.

Finetunes
1 models