TigerKay/magidonia-24b-lumia-cot

Hugging Face
TEXT GENERATIONConcurrency Cost:2Model Size:24BQuant:FP8Ctx Length:32kPublished:Mar 16, 2026License:apache-2.0Architecture:Transformer Open Weights Warm

TigerKay's Magidonia-24B-Lumia-CoT is a 24 billion parameter causal language model, fine-tuned using SaRA (Sparse Retraining Architecture) from TheDrummer's Magidonia-24B-v4.3. This model is specifically designed to produce detailed chain-of-thought reasoning before generating roleplay responses, following a 12-step framework. It excels at structured reasoning for conversational AI, particularly in roleplaying scenarios, and supports a 32K context length.

Loading preview...

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.