CrucibleLab/M3.2-24B-Loki-V1.3
CrucibleLab's M3.2-24B-Loki-V1.3 is a 24 billion parameter Mistral 3.2 finetuned model with a 32768 token context length. It is specifically trained on a 370 million token dataset curated for high-quality role-playing across diverse genres including fantasy, anime, sci-fi, and grimdark. This model excels at generating immersive storytelling experiences, leveraging its specialized dataset and rigorous multi-step data processing pipeline.
Loading preview...
Model Overview
M3.2-24B-Loki-V1.3 is a 24 billion parameter model developed by CrucibleLab, built upon the Mistral 3.2 architecture. It features a substantial 32768 token context window, enabling extensive and coherent conversational flows. The model's core strength lies in its specialized training on a 370 million token dataset meticulously curated for high-quality role-playing scenarios.
Key Capabilities
- Immersive Role-Playing: Designed to generate rich and detailed narratives across a wide array of genres, including:
- Fantasy (High Fantasy, Urban Fantasy)
- Anime & Manga
- Sci-Fi (Cyberpunk, Space Opera)
- ERP (Erotic Role-Play)
- Grimdark & Post-Apocalyptic
- Specialized Dataset: Trained on a unique dataset derived from well-established worlds and lore, ensuring deep thematic knowledge.
- Rigorous Data Processing: Utilizes a multi-step pipeline for data cleaning and enhancement, including unicode normalization, OOC misattribution fixes, and content filtering, to ensure a high-quality training base.
Recommended Usage
This model is particularly well-suited for applications requiring detailed and creative narrative generation, especially in interactive storytelling and role-playing contexts. It is recommended to use the standard Mistral template for optimal performance. CrucibleLab also provides several specialized system prompts (e.g., Hamon, Shingane, Kesshin, Kamae TTRPG) designed to guide the model for specific storytelling styles and character interactions. Recommended sampler settings include a static temperature of 0.7-1.0, Min P of 0.02-0.03, and specific DRY settings (Multiplier: 0.8, Base: 1.75, Length: 4-6) to fine-tune output creativity and coherence.