Shizuku AI unveils duplex model, advancing bidirectional AI companion development

Japanese AI laboratory Shizuku announced the release of its Shizuku-Duplex model. This new system is positioned as the second full-duplex model supporting intelligent Japanese dialogue, following GPT-Live, notable because it initiates thought processes without requiring a complete prompt. In Japanese language performance, Shizuku-Duplex achieved the highest conversational naturalness among the six tested models, ranking just behind OpenAI's GPT-Live in response quality. The architecture is split into a front-end interaction model and a background model capable of asynchronous complex request handling, incorporating context compression every five minutes. The model operates on an 80ms interaction frame and supports three output modalities: audio, text (model monologue), and actions. Shizuku AI plans to leverage this technology to develop its companion AI, Shizuku, which is designed to constantly listen, respond to user emotions, and express its own feelings openly.