QINGYI-KDA-0.6B: first successful Qwen3-to-KDA linearization (3:1 hybrid, GQA+RoPE kept layers) + persona alignment. Smallest trained KDA model, converted end-to-end on a consumer 8GB GPU.