llmfan46/Qwen3.5-27B-ultra-uncensored-heretic-v1
llmfan46/Qwen3.5-27B-ultra-uncensored-heretic-v1 is a 27 billion parameter Qwen3.5 model, developed by llmfan46, that has been decensored using the Heretic v1.2.0 method. This model significantly reduces refusals by 97% (3/100 vs 95/100 for the original) while maintaining high quality with a low KL divergence of 0.0301. It is optimized for applications requiring less restrictive content generation and robust performance across various language and vision tasks, including complex reasoning and coding.
Loading preview...
Overview
This model, llmfan46/Qwen3.5-27B-ultra-uncensored-heretic-v1, is a 27 billion parameter variant of the Qwen3.5 architecture, specifically modified by llmfan46 to be uncensored. It achieves a remarkable 97% reduction in refusals (3/100 compared to 95/100 in the original Qwen3.5-27B) while preserving model quality, indicated by a low KL divergence of 0.0301. The decensoring process utilized the Heretic v1.2.0 method, employing Magnitude-Preserving Orthogonal Ablation (MPOA) and Self-Organizing Map Abliteration (SOMA) on specific attention and MLP components.
Key Capabilities
- Significantly Reduced Refusals: Offers a highly uncensored output, ideal for use cases where content restrictions are undesirable.
- Preserved Performance: Maintains the strong capabilities of the base Qwen3.5-27B model across language, vision, and reasoning tasks, as evidenced by minimal impact on PIQA and MMLU scores.
- Multimodal Understanding: Inherits Qwen3.5's unified vision-language foundation, supporting complex visual understanding and multimodal reasoning.
- Long Context Handling: Features a native context length of 262,144 tokens, extensible up to 1,010,000 tokens, beneficial for intricate tasks.
Good for
- Applications requiring unrestricted content generation and creative freedom.
- Developers needing a powerful 27B parameter model with robust language and multimodal capabilities without inherent refusal mechanisms.
- Research into model safety and bias mitigation techniques, particularly in understanding the impact of decensoring methods.