DavidAU/Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinking

Model Overview

DavidAU/Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinking is a 21 billion parameter large language model based on the Qwen 3.6 architecture. It was created by taking an initial 27 billion parameter Qwen 3.6 'Heretic' model, reducing its size to 21 billion parameters, and then fine-tuning it across four datasets using Unsloth. A key characteristic of this model is its completely uncensored nature, achieved through the 'Heretic' modification by P-E-W, specifically implemented by 'trohrbaugh'.

Key Characteristics

• Architecture: Qwen 3.6 base, with a 21 billion parameter count after reduction from 27B.
• Uncensored: Explicitly designed for unrestricted content generation via the 'Heretic' modification.
• Fine-tuning: Utilizes Unsloth for fine-tuning across 4 datasets, with specific tuning to unify the model's layer structure post-reduction.
• Context Length: Supports a context length of 32768 tokens.

Performance & Considerations

Benchmarks provided by Nightmedia show some performance losses compared to the original 27B Qwen 3.6 model, particularly in 'thinking mode' versus 'instruct mode'. For example, in instruct mode, the 21B model scores 0.602 on arc-c compared to the 27B's 0.647. Users should be aware that the model may require additional tuning for both general and specific use cases to achieve desired performance levels. Its uncensored nature makes it suitable for applications where content filtering is not desired or where creative freedom is paramount.