DavidAU/Qwen3.6-27B-Heretic2-Uncensored-Finetune-Thinking

VISIONConcurrency Cost:2Model Size:27BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Apr 28, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

DavidAU/Qwen3.6-27B-Heretic2-Uncensored-Finetune-Thinking is a 27 billion parameter Qwen3.6-based language model, fine-tuned for uncensored responses and enhanced reasoning. It features a 32768 token context length and is specifically optimized to reduce refusal rates compared to its base model. This model is designed for applications requiring open-ended generation and improved performance across various benchmarks.

Loading preview...

Overview

DavidAU/Qwen3.6-27B-Heretic2-Uncensored-Finetune-Thinking is a 27 billion parameter model built upon the Qwen3.6 architecture, specifically fine-tuned to be uncensored and to enhance its core capabilities. This version, Heretic2, utilizes a base with an inherently lower refusal rate, which was further improved through a light fine-tuning process. The primary goal of this fine-tuning was to solidify the uncensored nature and boost critical metrics without heavy modifications.

Key Capabilities & Differentiators

  • Uncensored Generation: Explicitly designed for freedom in responses, with a significantly reduced refusal rate (4/100) compared to the original Qwen3.6-27B (99/100).
  • Enhanced Performance: Benchmarks indicate improved scores across various tasks (arc-c, arc/e, boolq, hswag, obkqa, piqa, wino) compared to both the base Qwen3.6-27B and the untuned Heretic version.
  • Low KL Divergence: Achieves a KL divergence of 0.0469 against the original model, indicating that while uncensored, its generation patterns remain closely aligned with the original model's performance characteristics.
  • Optimized for Open-Ended Use: The fine-tuning process aimed at minor, critical fixes to ensure robust uncensored output while maintaining high performance.

Good For

  • Applications requiring unrestricted and creative text generation.
  • Use cases where low refusal rates are critical for user experience.
  • Scenarios demanding improved reasoning and general knowledge performance over standard Qwen3.6 models.