Nimbz/Schattenblume-31B
Schattenblume-31B is a 31 billion parameter instruction-tuned language model developed by Nimbz, built on the Gemma 4 architecture. This merge model is specifically designed for roleplay, creative writing, and character adherence, leveraging a unique blend of three donor models to reduce 'Gemma-isms' and enhance prose quality. It excels at maintaining in-character responses and providing diverse narrative options, making it suitable for interactive storytelling and detailed character interactions.
Loading preview...
Schattenblume-31B: A Roleplay-Focused Gemma 4 Merge
Schattenblume-31B is a 31 billion parameter language model developed by Nimbz, created by merging three distinct models on a Gemma 4 base. This model is primarily engineered for roleplay and creative writing, with a strong emphasis on in-character adherence and diverse narrative generation.
Key Capabilities & Design Philosophy
- Enhanced Roleplay: Designed to play characters closer to their written descriptions, pushing back when appropriate, rather than simply agreeing to all user prompts.
- Reduced 'Gemma-isms': Integrates
scotoma-2to mitigate common stylistic quirks found in Gemma models, improving output quality and naturalness. - Superior Prose & Imagery: Leverages
Giftige-Blume-v1to enhance the model's ability to generate rich, descriptive prose and vivid imagery. - Swipe Variety & Entity Tracking: Incorporates
MeroMero-v2to ensure a wide range of narrative options and consistent tracking of entities within a conversation. - Layer-Specific Merging: Unlike typical merges, Schattenblume-31B uses a
della_linearmerge method with carefully weighted layers, targeting specific strengths of each donor model to optimize for character work, prose, and output correction.
When to Use This Model
Schattenblume-31B is ideal for applications requiring:
- Detailed and immersive roleplaying scenarios.
- Creative writing tasks where rich descriptions and strong character voices are paramount.
- Interactive storytelling that benefits from a model capable of maintaining consistent character traits and narrative flow.
Users are encouraged to provide a detailed scene or context to fully leverage the model's capabilities. For optimal performance, specific sampler settings are recommended, including a Temperature of 0.8-1.1 and Min-P of 0.05-0.15, with repetition penalty turned off. The model also supports a "<|think|>" token at the start of the system prompt for enhanced reasoning in token-heavy scenarios.