llmfan46/Qwen3.5-9B-ultra-uncensored-heretic-v1
The llmfan46/Qwen3.5-9B-ultra-uncensored-heretic-v1 is a 9 billion parameter causal language model, based on the Qwen3.5-9B architecture, with a 32768 token context length. This model has been decensored using Heretic v1.2.0 with Magnitude-Preserving Orthogonal Ablation (MPOA) and Self-Organizing Map Abliteration (SOMA) techniques. It significantly reduces content refusals while largely preserving the original model's capabilities, making it suitable for applications requiring less restrictive content generation.
Loading preview...
Model Overview
llmfan46/Qwen3.5-9B-ultra-uncensored-heretic-v1 is a 9 billion parameter model derived from the Qwen3.5-9B architecture, featuring a 32768 token context length. This version has been specifically modified using the Heretic v1.2.0 framework, employing Magnitude-Preserving Orthogonal Ablation (MPOA) and Self-Organizing Map Abliteration (SOMA) to reduce content censorship.
Key Differentiators
- Decensored Output: Achieves a significantly lower refusal rate (2/100) compared to the original Qwen3.5-9B (86/100), enabling less restricted content generation.
- Capability Preservation: Maintains strong performance across various benchmarks, with a KL divergence of 0.1085 from the original model, indicating minimal loss in core abilities like reasoning and common-sense understanding (e.g., PIQA scores remain close to the original).
- Multimodal Foundation: Inherits Qwen3.5's unified vision-language foundation, efficient hybrid architecture, and scalable RL generalization, supporting expanded linguistic coverage across 201 languages and dialects.
- Extended Context: Natively supports a context length of 262,144 tokens, extensible up to 1,010,000 tokens using YaRN scaling techniques, beneficial for processing ultra-long texts.
Ideal Use Cases
- Applications requiring less restrictive or uncensored text generation.
- Tasks benefiting from multimodal input (image/video) and long context understanding.
- Agentic applications leveraging its tool-calling capabilities, especially with frameworks like Qwen-Agent or Qwen Code.