Magidonia 24B — Lumia CoT Overview
Magidonia 24B — Lumia CoT is a specialized 24 billion parameter language model developed by TigerKay, based on TheDrummer's Magidonia-24B-v4.3. Its core innovation lies in its ability to generate a detailed <thinking> block prior to each response, guiding its output through a structured 12-step reasoning process. This makes it particularly adept at complex, multi-turn conversational tasks, especially in roleplaying.
Key Capabilities & Features
- Chain-of-Thought Reasoning: Integrates a unique 12-step reasoning framework (Lucid Loom) including scene analysis, character headspace, and content planning, before generating responses.
- SaRA Fine-tuning: Utilizes Sparse Retraining Architecture, modifying 15% of the base model's weights, trained on over 10,000 chain-of-thought traces from Bluemoon RP conversations.
- Context Length: Supports a substantial 32,768 token context window, suitable for extended interactions.
- Native Tag Support: Designed to work natively with
<thinking> </thinking> tags (Mistral Tekken format) for explicit reasoning.
Ideal Use Cases
- Advanced Roleplay: Optimized for generating coherent and contextually rich roleplay responses by first outlining its reasoning process.
- Structured Conversational AI: Suitable for applications requiring transparent, step-by-step reasoning in dialogue generation.
- Interactive Storytelling: Can be leveraged for creating dynamic narratives where the AI's thought process is part of the experience.
For optimal performance, prefilling the assistant with <thinking> is recommended to initiate the reasoning pattern, improving first-message reliability.