Shizuku AI has released Shizuku-Duplex, a Japanese full-duplex dialogue model that listens and speaks at the same time, so an AI character can give short backchannels, start answering before a sentence ends, and stop when the user cuts in. It is built for companion characters, and the company is applying it to Shizuku, the character it has developed since a YouTube debut in 2023. Cascaded voice systems still wait to detect the end of a turn, then run recognition, a language model, and synthesis in sequence. Aside from OpenAI's GPT-Live, Shizuku AI reports this as the first full-duplex model capable of intelligent dialogue in Japanese. The model steps forward in 80 millisecond ticks, hearing the user and producing speech on every tick, while writing the next words as text slightly ahead of the audio. The same step can select a Live2D motion. When a request needs a tool such as web search, an interaction model hands it to a background model, keeps the conversation going, and later states the result in its own words. On the company's Japanese benchmark, the median time from the end of speech to a reply was 0.57 seconds, against 0.53 seconds for GPT-Live. Among full-duplex models it led on five of six turn-taking metrics and placed second to GPT-Live on reply quality, ahead of J-Moshi and LLM-jp-Moshi. A launch demo shows the character taking a correction in the middle of a plan, including a small gift budget and a dislike of flowers, and checking a live fact without leaving a silent gap. Founder and CEO Akio Kodaira started the Shizuku character in 2023 while completing a Ph. D. at UC Berkeley.