Gryphe/Qwen3.6-35B-A3B-StyleTune
Gryphe/Qwen3.6-35B-A3B-StyleTune is a 35.1 billion parameter Qwen3.6-based language model developed by Gryphe, specifically fine-tuned for narrative writing style. This model achieves a 46.9% reduction in clichés and a significantly altered vocabulary by targeting only the lm_head output projection, preserving all original reasoning and knowledge capabilities. It is optimized for generating creative and distinct narrative text while maintaining the base model's core intelligence.
Loading preview...
Overview
Gryphe/Qwen3.6-35B-A3B-StyleTune is a specialized 35.1 billion parameter model built upon the Qwen3.6-35B-A3B architecture. Unlike traditional fine-tuning that modifies many layers, this model undergoes a "style tune" by training only a single tensor: the lm_head output projection. This targeted approach dramatically reduces VRAM requirements and training time, allowing for rapid style modification on consumer hardware while leaving the core model's capabilities entirely intact.
Key Capabilities
- Distinct Writing Style: Achieves a 46.9% reduction in clichés and an 80.8% different trigram vocabulary compared to the base instruct model, resulting in much less repetitive and more unique narrative outputs.
- Preserved Core Intelligence: All of Qwen3.6's original reasoning, world knowledge, instruction following, and language understanding capabilities remain fully functional, as only the output style layer is modified.
- Efficient Fine-tuning: Demonstrates that significant stylistic changes can be achieved with minimal computational resources by focusing on the
lm_headtensor.
When to Use This Model
- Narrative Generation: Ideal for applications requiring creative writing, storytelling, or roleplay where a unique and less cliché-ridden voice is desired.
- Maintaining Base Model Strengths: Suitable for tasks where the robust reasoning and knowledge of Qwen3.6 are essential, but a specific stylistic output is also required.
- Experimentation with Style: Useful for developers interested in exploring how targeted, single-tensor fine-tuning can alter model output without compromising core functionality.