llmfan46/Qwen3.5-27B-uncensored-heretic-v2-Native-MTP-Preserved

VISIONConcurrency Cost:2Model Size:27BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:May 24, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

The llmfan46/Qwen3.5-27B-uncensored-heretic-v2-Native-MTP-Preserved is a 27 billion parameter Qwen3.5-based causal language model, developed by llmfan46, with a native context length of 32768 tokens. This model is a decensored version of Qwen/Qwen3.5-27B, achieved using the Heretic v1.3.0 method with Magnitude-Preserving Orthogonal Ablation (MPOA) to significantly reduce refusals by 89% while maintaining original model quality (0.0308 KL divergence). It is optimized for applications requiring less content restriction and robust performance across various language and multimodal tasks.

Loading preview...

Model Overview

This model, llmfan46/Qwen3.5-27B-uncensored-heretic-v2-Native-MTP-Preserved, is a 27 billion parameter variant of the Qwen3.5 base model, developed by llmfan46. It has been specifically modified using the Heretic v1.3.0 method with Magnitude-Preserving Orthogonal Ablation (MPOA) to achieve a significant reduction in content refusals.

Key Differentiators

  • Decensored Performance: Achieves an 89% reduction in refusals (10/100 vs. 95/100 for the original model) while preserving model quality with a low KL divergence of 0.0308.
  • Native MTP Preservation: All 15 Multi-Token Prediction (MTP) components from the original Qwen3.5-27B model are preserved, ensuring robust performance.
  • Multimodal Capabilities: Inherits Qwen3.5's unified vision-language foundation, supporting image and video inputs, and excelling in tasks like VQA, text recognition, and spatial intelligence.
  • Extended Context Length: Natively supports a context length of 262,144 tokens, extensible up to 1,010,000 tokens using YaRN scaling techniques.

Use Cases

This model is particularly well-suited for applications where reduced content restrictions are desired, such as creative writing, open-ended dialogue, and research requiring unfiltered information. Its strong performance in instruction following, coding, and agentic tasks, combined with multimodal capabilities, makes it versatile for complex AI applications.

Finetunes
5 models