Nimbz/Versipellis-31B
Nimbz/Versipellis-31B is a 31 billion parameter Gemma 4-based merge model designed for highly steerable, uncensored, and in-character conversational roleplay. It excels at maintaining character consistency and long-context detail, treating character descriptions as its primary directive. The model focuses on generating natural, dialogue-forward prose with short, physical action descriptions, making it ideal for immersive interactive storytelling.
Loading preview...
Versipellis-31B: The Shapeshifting Conversationalist
Versipellis-31B is a 31 billion parameter model built on the Gemma 4 architecture, specifically engineered to be an authentic and highly steerable conversational partner. Unlike many models that prioritize user-adherence, Versipellis-31B strictly follows character descriptions, allowing it to embody diverse personas without breaking character or resorting to disclaimers. It maintains a clean distinction between the model's character, the user, and any introduced third parties.
Key Capabilities:
- Deep Character Immersion: Treats character descriptions as absolute law, enabling genuine in-character responses.
- Natural Dialogue & Prose: Generates short, dialogue-forward turns with concise physical action descriptions, avoiding verbose narration.
- Long-Context Coherence: Reliably tracks details over extended conversations (e.g., clothing, spatial state, past events) due to shielded global-attention layers during merging.
- High Steerability: Responds well to mid-chat steering for tone shifts or persona changes, settling back naturally afterward.
- Uncensored Roleplay: Designed for uncensored, in-character roleplay without inherent agreeability.
Merge Architecture:
The model was developed through a six-phase merge process using mergekit's gemma-4-support fork. This involved blending various models to create distinct "Spirit" (creative prose), "Personality" (conversational core), "Voice" (vocabulary and prose style), and "Mind" (reasoning) components, which were then combined to form the final "Soul" and ultimately, Versipellis. Special attention was given to protecting Gemma 4's long-context global-attention layers to ensure stability.
Recommended Usage:
- System Prompt: Utilize the provided system prompt, which includes options for
[SHORT]or[NARRATE]action styles, to guide the model's output. - Sampling: Recommended sampler settings include Temperature 0.7-0.9, Top-P 0.98, and Min-P 0.05-0.1, with repetition penalty set to off.
- Ongoing Chats: For longer conversations, a low-depth reminder (e.g.,
"[Stay in conversation: speak in {{char}}'s voice, narrate only {{char}}'s own physical actions - never the scene. Lead with dialogue.]") can help maintain desired output style.