SubMaroon/Boulesis-v2.1-26B-A4B
Boulesis-v2.1-26B-A4B is a 26 billion parameter composite Gemma 4 model developed by SubMaroon, utilizing QK task arithmetic and fused LoRA. This version is specifically designed to enhance character understanding, context memory, and prose diversification for complex roleplay scenarios. It aims to provide a more decisive narrative while retaining core intelligence and knowledge, making it suitable for intricate interactive storytelling.
Loading preview...
Overview
Boulesis-v2.1-26B-A4B is a 26 billion parameter model based on the Gemma 4 architecture, developed by SubMaroon. It is a composite model created using QK task arithmetic and fused LoRA techniques. The primary goal of this version is to refine the model's ability to understand and retain character information and context, leading to more immersive and coherent roleplay experiences.
Key Enhancements in v2.1
- Enhanced Character Grasp: This version demonstrates the best understanding of characters among the Boulesis iterations.
- Superior Context Memory: It excels at remembering and utilizing context throughout long roleplay sessions, integrating lore and facts organically.
- Balanced Prose: Offers a more measured tone compared to v1, while maintaining a dynamic plot and driving character actions.
- Optimized for Roleplay: Specifically tuned to diversify prose and make the model more decisive in narrative progression.
Differences from Previous Versions
Boulesis-v2.1 builds upon v1 and v2 with specific adjustments to QK donor, alpha values, LoRA layers, and bake scale. Notably, v2.1 maintains the English-only training data introduced in v2, focusing on refining the model's core capabilities for complex interactive narratives. It is considered by the developer to be the best version for intricate roleplay sessions involving multiple characters.
Recommended Usage
For optimal performance, the developer recommends specific settings:
- 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 interactions, especially regarding thinking enablement and sampler settings.