mirxa2/Qwen3.6-9B-Thinking-Sweet-Madness
mirxa2/Qwen3.6-9B-Thinking-Sweet-Madness is a 27 billion parameter Qwen 3.6 model, uncensored and then compressed to 9 billion parameters with 16 layers, fine-tuned across multiple datasets. This model is specifically designed for highly creative, uncensored, and "mad" text generation, excelling in imaginative and unconventional outputs. It retains image/video training capabilities from its larger base model and supports a 256k context length.
Loading preview...
mirxa2/Qwen3.6-9B-Thinking-Sweet-Madness: Creative Madness Unleashed
This model is a unique iteration of the Qwen 3.6 architecture, initially a 27 billion parameter model that underwent a significant transformation. It was first uncensored using the Heretic method by P-E-W (heretic'ing by "trohrbaugh"), then "shrunk" to 9 billion parameters via a modified Mergekit process, reducing its layers from 64 to 16. The resulting 9B model was subsequently fine-tuned on local hardware using Unsloth across six datasets in two stages, with specific tuning to unify its new layer structure.
Key Capabilities & Characteristics
- Uncensored & Creative: Designed for highly imaginative, unconventional, and uncensored text generation, often described as "creative madness gold."
- Reduced Size, Retained Power: Despite being compressed from 27B to 9B parameters and 16 layers, it retains core functionalities, including image/video training capabilities from the full 27B model.
- Fast Inference: Achieves approximately 200 tokens/second on a 5090 GPU (Q4KS quantization), partly due to its reduced layer count.
- Large Context Window: Supports a substantial 256k context length.
- "Thinking" Process: Example generations demonstrate an internal "thinking" process, providing insight into its generation strategy.
Important Considerations & Use Cases
This model is explicitly noted as being "mad" and should not be relied upon for truth or productive capacity without further tuning. It is a "lab" model in its current state, best suited for:
- Creative Writing & Roleplay: Excels in generating unique, imaginative, and unrestricted narratives.
- Exploratory AI Research: Ideal for experimenting with unconventional model behaviors and outputs.
- Further Fine-tuning: Requires additional tuning (25-50k samples minimum) for specific or general use cases to reach its full potential, as its performance may be surpassed by a standard Qwen 3.5 9B until fully tuned. Users should be aware that knowledge or skills from the original 27B model may be missing due to the compression process.