Nimbz/Gemma-4-Dark-Gemistry-31B
Nimbz/Gemma-4-Dark-Gemistry-31B is a 31 billion parameter Gemma 4-based merge model with a 32768 token context length, specifically designed for uncensored, gritty, and creative roleplay. It integrates multiple fine-tunes to enhance swipe variety, character agency, and detailed anatomical descriptions, while maintaining long-context coherence. The model features a unique untied lm_head for a distinct prose style, making it ideal for verbose and unflinching narrative generation.
Loading preview...
Dark-Gemistry: Uncensored Roleplay with a Distinct Voice
Nimbz/Gemma-4-Dark-Gemistry-31B is a 31 billion parameter Gemma 4-based merge model, engineered for rich, uncensored, and gritty roleplay experiences. It leverages a sophisticated multi-phase merging process to combine the strengths of several specialized fine-tunes, resulting in a model that excels in creative narrative generation.
Key Capabilities & Features
- Uncensored & Gritty Roleplay: Built on a "heretic-ara" backbone, it strips refusals and embraces explicit, unflinching content.
- Enhanced Creativity & Intelligence: Integrates "Gemsicle-31B" for strong roleplay intelligence and creative depth, and "MeroMero" for diverse swipe variety and character agency.
- Distinct Prose Style: Features a custom, untied
lm_headfrom "StyleTune" that bakes in a unique voice, eliminating the need for explicit prose style instructions in prompts. - Long-Context Coherence: Gemma 4's global-attention layers are shielded, preserving 256K long-context coherence across extended chats.
- Detailed & Paced Narratives: Excels at tracking long-context details, physical states, and generating well-paced, anatomically detailed scenes.
- Image Prompt Generation: Demonstrates strong creativity in image prompt generation while maintaining adherence to given structures.
When to Use This Model
Dark-Gemistry is particularly well-suited for:
- Creative Writing & Roleplay: Ideal for generating verbose, detailed, and explicit roleplay scenarios.
- Narrative Generation: When a distinct, consistent narrative voice is desired without explicit prompting.
- Applications Requiring Unflinching Content: For use cases where uncensored and gritty themes are central to the output.
Sampler recommendations suggest a temperature range of 0.75-1.0, with repetition penalties generally turned off to allow the lm_head's inherent style to flourish. The model also supports an optional system prompt with "thinking" capabilities and adjustable action styles (short or narrative).