The model generates a <thinking> block before each response, following the Lucid Loom 12-step reasoning framework: scene analysis, character headspace, objectives tracking, content planning, and more.

Training

Method: SaRA (Sparse Retraining Architecture) — 15% of weights (3.3B / 22B)
Dataset: 10,374 chain-of-thought reasoning traces over 1,041 Bluemoon RP conversations
Format: Multi-turn windows, 3 thinking turns per window, 32K context
Base model: Magidonia-24B-v4.3 (Mistral v3, Tekken v7)
Training: 1000 steps, LR 5e-5, cosine schedule
Hardware: NVIDIA B200

Quantizations (GGUF with imatrix)

Quant	Size	Min VRAM (32K ctx)
Q4_K_M	~14 GB	~19 GB (4090)
Q5_K_M	~16 GB	~22 GB (4090)
Q6_K	~19 GB	~25 GB
Q8_0	~25 GB	~31 GB

Usage

Works with SillyTavern, KoboldCpp, llama.cpp, LM Studio.

Uses <thinking> </thinking> tags natively (Mistral Tekken format).

For best results, prefill assistant with <thinking> to ensure reasoning starts.

Status

First-message reliability ~25% without prefill, improves with prefill
Once pattern starts, maintains consistently
DPO alignment planned to improve reliability

TigerKay/magidonia-24b-lumia-cot

Magidonia 24B — Lumia CoT

What it does

Training

Quantizations (GGUF with imatrix)

Usage

Status

TigerKay/
magidonia-24b-lumia-cot