Hugging Face Overview
CrucibleLab's M3.2-24B-Loki-V2 is a 24 billion parameter model specifically fine-tuned for deeply immersive roleplay and narrative experiences. It prioritizes immersion over compliance, aiming to create models that act as creative partners rather than restrictive assistants. The model is designed to behave like Dungeonmasters, Story authors, or Creative narrators, focusing on tone, voice, continuity, and world fidelity for consistent long-form storytelling.
Key Differentiators
- Clear narrative authority: Prevents role confusion between narrator and player.
- Strong immersion: Focuses on stories, not generic assistant responses.
- Consistent long-form storytelling: Maintains voice and continuity over extended sessions.
- Flexible but controlled creativity: Expands naturally within defined boundaries.
- Universe-conditioned writing: Trained on various universes (e.g., Attack on Titan, Dragon Age, Warhammer 40,000) to match tone, themes, and terminology, but is also highly versatile for original settings.
Narrative Modes & Tone Selection
The model supports distinct narrative modes chosen via the system prompt:
- Dungeonmaster (DM): Controls world and NPCs, enforces consequences.
- Storywriter: Follows OOC instructions, maintains lore and character voice.
- Creative Storyteller: Adapts archetypes and expands scenes creatively.
Users can also select a tone: Neutral, Dark, or Mature, to shape the story's emotional weight. The model also incorporates advanced tagging for ALT_ANATOMY and other specific content, allowing for highly customized and flexible narrative generation.
Training & Future Horizons
Loki-v2 is a 1 Epoch fine-tune. The developers aim to bring the Loki dataset to much larger architectures (70B+) to unlock deeper nuance and longer context retention, contingent on securing necessary compute resources.