Kakao unveiled its Kanana Space AI assistant for KakaoTalk on 13 October, with a closed beta planned for December, Yonhap reported. The company presented the service at its annual developers’ conference at its AI campus in Yongin, Gyeonggi Province.
Key points
- Kakao plans a closed beta of Kanana Space in December 2026 and an official launch in the first quarter of 2027.
- The assistant is designed to act on KakaoTalk requests by voice or text, including messages, schedules and app settings.
- Kakao also presented Kanvas, which lets users create PC agents, and Kanana-O, a model it plans to release as open source.
Kanana Space moves into KakaoTalk tasks
Kakao opened applications for the Kanana Space trial on 13 October and plans to select participants for the December beta, The Asia Business Daily reported. An official launch is scheduled for the first quarter of 2027. The trial puts the service on a timetable beyond the demonstration Kakao gave at the conference.
In that demonstration, Kakao design lead Kim Daenyeon pulled down the KakaoTalk screen with two fingers to summon the assistant, represented by the character Ryan. Kim asked it to mark unread messages in a folder as read, and the messages were marked accordingly. Kakao says users will be able to give requests by voice or text, including sending messages, organising schedules, searching the web and changing KakaoTalk settings.
Another demonstration showed the assistant recommending exercise clothes using the weather and a user’s workout data, or arranging a schedule from a user’s KakaoTalk message history, Yonhap reported. Those examples depend on the service understanding information already held in, or supplied to, Kakao’s products rather than responding only to an isolated prompt.
Kakao plans to extend Kanana Space to ordering goods, sending gifts, booking services and transferring money through spoken requests. The Asia Business Daily described an example in which a request to order bottled water would prompt the assistant to select a previously purchased item and apply available coupons. The company presented this as a direction for the service, rather than a transaction completed in the demonstration.
Chung said Kakao intended to move from a graphical interface towards one that interprets the speaker’s language, purpose and context. She also said the measure of a generative AI system was shifting from the quality of its answers towards its ability to complete requested tasks. For KakaoTalk, that makes carrying out an instruction inside the app central to the proposed service.
Kanana models underpin the assistant
Kakao presented its Kanana foundation-model work alongside the assistant. An executive responsible for the models said the company sought to handle requests without using its largest model every time. Kakao applied Multi-Teacher On-Policy Distillation, a post-training method, to maintain the quality of smaller language models, the executive said.
The company also introduced Kanana-O, which it says can process text, images and voice together. Kakao developed the model’s encoders and decoders in-house and plans to release it as open source, The Asia Business Daily reported. Noh Byungseok, who leads Kakao’s foundation-model development, said the company had worked on how the models are made smaller, how much computation they use and how they serve requests.
Kakao also cited a Korean tokenizer and quantisation-aware training among its methods for improving inference efficiency. The model work sits behind the assistant’s planned mix of spoken requests, personal context and actions across KakaoTalk.
Kanvas takes Kakao agents onto PCs
Kakao separately unveiled Kanvas, a service for creating and running agents on a user’s PC. Users describe the agent they want, and Kanvas connects tools and capabilities for the task, according to the company. Kakao says those agents will act within permissions set by users, who can review their work at key stages rather than direct each step.
Kakao plans a small closed beta for Kanvas as early as 2026 and is preparing a way for users to share agents they create. The company is also developing Project Mosaic, which it says will help agents use relevant information from a user’s past activity. Chief technology officer Song Jaeha said agents should receive only the information needed for an authorised task.
Kakao plans to add a space in KakaoTalk in the first half of 2027 where users can connect a group of agents tailored to them, including agents made with Kanvas and those built using other frameworks.
