DavidAU/Qwen3.5-27B-HERETIC-Polaris-Advanced-Thinking-Alpha-uncensored

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

The DavidAU/Qwen3.5-27B-HERETIC-Polaris-Advanced-Thinking-Alpha-uncensored is a 27 billion parameter Qwen 3.5 dense model fine-tuned by DavidAU using the POLARIS distill dataset. This model features altered reasoning/thinking blocks and is a "HERETIC" model, designed to follow instructions without refusal. It supports vision inputs and maintains strong benchmarks while offering reduced safety alignment.

Loading preview...

Model Overview

This model, DavidAU/Qwen3.5-27B-HERETIC-Polaris-Advanced-Thinking-Alpha-uncensored, is a 27 billion parameter Qwen 3.5 dense model fine-tuned by DavidAU. It incorporates a unique "HERETIC" training approach, significantly reducing safety alignment and refusals (14/100 compared to 94/100 for the original Qwen3.5-27B). The fine-tuning process also altered its reasoning/thinking blocks and their size, aiming for enhanced instruction following.

Key Capabilities

  • Uncensored Instruction Following: Designed to execute commands without refusal, offering greater flexibility for developers.
  • Vision-Language Support: Tested and confirmed working with image inputs, leveraging the base Qwen3.5's unified vision-language foundation.
  • Altered Reasoning Architecture: Features modified internal reasoning mechanisms for potentially different response generation.
  • Strong Benchmark Retention: Despite modifications, efforts were made to preserve the strong performance of the base Qwen3.5 model across various benchmarks.
  • Multilingual Support: Inherits Qwen3.5's expanded support for 201 languages and dialects.
  • Long Context: Supports a native context length of 262,144 tokens, extensible up to 1,010,000 tokens with YaRN scaling.

Good for

  • Unrestricted Content Generation: Ideal for applications requiring responses without built-in safety filters or refusals.
  • Advanced Reasoning Tasks: Potentially beneficial for tasks where modified thinking blocks could offer novel approaches.
  • Multimodal Applications: Suitable for tasks involving both text and image inputs.
  • Developers requiring maximum control: Offers a model that prioritizes user instructions over inherent safety alignments.