Kakao Develops On-Device Multimodal Model

Kakao is developing and refining Kanana-o, an on-device multimodal AI model that can run on smartphones and other devices. The omni model processes text, images, and audio, and Kakao has secured a version that runs directly on a device. The company expects to apply it to KakaoTalk AI services such as Kanana Space.

Kanana Space appears in KakaoTalk when users swipe down with two fingers and lets them send requests to the character Ryan by voice or text. Kakao plans to begin a closed beta test (CBT) in December; at first, the service will prioritize the server-based model. Over the longer term, Kakao plans to handle simple tasks on smartphones and assign complex tasks to frontier models to improve efficiency.

On-device AI runs models directly on a device, reducing reliance on servers and improving response speed. Kakao says it can also help privacy and security by limiting how much user information is sent to external servers. Noh Byung-seok, Kakao's Unified Foundation Model performance lead, said on the 13th at if(kakao)26 that handling requests on a smartphone before sending them to a server can speed up responses and reduce the amount of sensitive information sent outside the device.

Kakao is also developing encoders to support its multimodal AI technology. On the 21st, it released its in-house vision encoder as open source and said it plans to release its audio encoder within the year. Kakao said it would use its understanding of Korean and user context to improve model performance and efficiency, advance foundational technology for AI agent services, and share its research results.

Kakao is developing and refining Kanana-o, an on-device multimodal AI model that can run on smartphones and other devices. The omni model processes text, images, and audio, and Kakao has secured a version that runs directly on a device. The company expects to apply it to KakaoTalk AI services such as Kanana Space.

Kanana Space appears in KakaoTalk when users swipe down with two fingers and lets them send requests to the character Ryan by voice or text. Kakao plans to begin a closed beta test (CBT) in December; at first, the service will prioritize the server-based model. Over the longer term, Kakao plans to handle simple tasks on smartphones and assign complex tasks to frontier models to improve efficiency.

On-device AI runs models directly on a device, reducing reliance on servers and improving response speed. Kakao says it can also help privacy and security by limiting how much user information is sent to external servers. Noh Byung-seok, Kakao's Unified Foundation Model performance lead, said on the 13th at if(kakao)26 that handling requests on a smartphone before sending them to a server can speed up responses and reduce the amount of sensitive information sent outside the device.

Kakao is also developing encoders to support its multimodal AI technology. On the 21st, it released its in-house vision encoder as open source and said it plans to release its audio encoder within the year. Kakao said it would use its understanding of Korean and user context to improve model performance and efficiency, advance foundational technology for AI agent services, and share its research results.