DavidAU/Qwen3.6-9B-Heretic-Uncensored-Thinking-Sweet-Madness

VISIONConcurrency Cost:2Model Size:27BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:May 20, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

DavidAU/Qwen3.6-9B-Heretic-Uncensored-Thinking-Sweet-Madness is a 27 billion parameter Qwen 3.6 model, reduced to 9 billion parameters and 16 layers via a modified Mergekit process. This model is fully uncensored and fine-tuned for highly creative, unconventional, and 'mad' text generation, including detailed 'thinking' processes. It excels in generating unique, imaginative, and often unpredictable content, making it suitable for experimental creative writing and exploring non-standard AI outputs.

Loading preview...

Overview

DavidAU/Qwen3.6-9B-Heretic-Uncensored-Thinking-Sweet-Madness is a unique 9 billion parameter model derived from the Qwen 3.6 27B architecture. It underwent a "Heretic" uncensoring process by P-E-W (heretic'ing by "trohrbaugh") and was subsequently "shrunk" to 9B parameters and 16 layers using a modified Mergekit. The model was then fine-tuned on local hardware via Unsloth across six datasets in two stages, with specific tuning to unify its new layer structure.

Key Capabilities

  • Uncensored Generation: Designed to be fully uncensored, allowing for unrestricted content creation.
  • Creative & Unconventional Output: Excels at generating highly imaginative, 'mad,' and unpredictable text, often including detailed 'thinking' processes.
  • Fast Inference: Achieves speeds of 200 tokens/second on a 5090 GPU (Q4KS quantization), partly due to its reduced 16-layer structure.
  • Multimodal Foundations: Retains intact image/video training and systems lifted from the full 27B model.
  • Extended Context: Supports a 256k context length.

Important Considerations

  • Experimental Nature: Described as "Sweet Madness," it's an experimental model that may require additional tuning for specific or general use cases.
  • Knowledge Gaps: Due to the unique compression method, some knowledge or skills from the original 27B model may be missing.
  • Performance vs. Qwen 3.5 9B: The model's performance may be surpassed by Qwen 3.5 9B until it is fully tuned with a minimum of 25-50k samples.

Good For

  • Experimental Creative Writing: Ideal for generating highly imaginative, non-linear, or 'mad' narratives and dialogues.
  • Exploring Unconventional AI Behavior: Useful for researchers or developers interested in pushing the boundaries of AI output and exploring uncensored responses.
  • Rapid Prototyping: Its fast inference speed makes it suitable for quick generation of creative content.

Settings Recommendations

  • Temperature: 1
  • Top P: 0.95
  • Min P: 0.05
  • Repetition Penalty: 1.05 (1.1 strongly suggested, especially for short prompts)
  • Context: Minimum 8k to 16k
  • Quantization: Q4KS minimum or IQ3_M Imatrix

Finetunes
1 models