SubMaroon/Boulesis-v2.1-26B-A4B-HeadTune
SubMaroon/Boulesis-v2.1-26B-A4B-HeadTune is a 26 billion parameter experimental composite Gemma 4 model, developed by SubMaroon, utilizing QK task arithmetic, fused LoRA, and a custom `lm_head`. This version is specifically tuned for a very literary and descriptive writing style, making it suitable for users who prefer highly ornate and expressive text generation in roleplay scenarios. It features a 32768 token context length, focusing on character grasp and context memory.
Loading preview...
Model Overview
Boulesis-v2.1-26B-A4B-HeadTune is an experimental 26 billion parameter model based on the Gemma 4 architecture, developed by SubMaroon. It incorporates QK task arithmetic, fused LoRA, and a custom lm_head trained on a unique dataset. This specific variant is characterized by its highly literary and descriptive writing style, employing numerous turns of phrase.
Key Characteristics & Differentiators
This model is part of the Boulesis series, with v2.1 being noted for its excellent grasp of characters and strong context memory in complex roleplay sessions. The "Headtune" variant further refines this by emphasizing a verbose, descriptive narrative. It is an experimental merge, with ongoing feedback collection to improve future iterations.
Recommended Usage
For optimal performance, specific sampler settings are recommended:
- Temperature:
1.0 - Top-K:
64 - Repetition Penalty:
1.05-1.1 - Top-P:
0.95
Users are encouraged to provide feedback on its performance in long sessions, scene logic, and character consistency, especially when using different sampler settings or with "thinking" enabled/disabled.