llmfan46/Qwen3.5-27B-ultra-uncensored-heretic-v2

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Mar 31, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

llmfan46/Qwen3.5-27B-ultra-uncensored-heretic-v2 is a 27 billion parameter language model based on the Qwen3.5 architecture, specifically decensored using the Heretic v1.2.0 tool with Arbitrary-Rank Ablation (ARA) method. This model significantly reduces refusals to 1/100 compared to the original's 95/100, while preserving model quality with a KL divergence of 0.0600. It maintains strong performance across various benchmarks including MMLU, PIQA, and coding tasks, making it suitable for applications requiring less restrictive content generation.

Loading preview...

Model Overview

llmfan46/Qwen3.5-27B-ultra-uncensored-heretic-v2 is a 27 billion parameter model derived from the Qwen/Qwen3.5-27B base, engineered for significantly reduced content refusals. It achieves a refusal rate of 1/100, a substantial improvement over the original model's 95/100, while maintaining a low KL divergence of 0.0600, indicating strong preservation of the base model's capabilities. This decensoring was performed using the Heretic v1.2.0 tool with the Arbitrary-Rank Ablation (ARA) method, specifically targeting attention output and object projection layers.

Key Capabilities

  • Reduced Refusals: Offers 99% fewer refusals, enabling more open-ended and less restricted content generation.
  • Capability Preservation: Benchmarks like PIQA and MMLU show minimal degradation in reasoning and general knowledge compared to the original Qwen3.5-27B.
  • Multimodal Foundation: Inherits Qwen3.5's unified vision-language foundation, supporting multimodal tokens for reasoning, coding, agents, and visual understanding.
  • Long Context: Supports a native context length of 262,144 tokens, extensible up to 1,010,000 tokens, beneficial for complex tasks.

Good For

  • Applications requiring highly uncensored or less restrictive text generation.
  • Tasks benefiting from a large context window and strong multimodal understanding.
  • Developers seeking a powerful 27B parameter model with robust general capabilities and reduced content filtering.

Finetunes
3 models