Kakao Hosts 'Kanana Scholar' Seminar with Academia to Discuss 11T Token Training Efficiency and Industry-Academia Collaboration

Kakao held the first seminar of its AI collaboration program 'Kanana Scholar' with academic experts at 'Kakao Camp Gangnam' near Samseong Station. Seven leading AI scholars from major universities including KAIST, Seoul National University, Yonsei University, and Hanyang University, along with Kakao AI leaders, attended the event to discuss securing technological sovereignty and industry-academia collaboration strategies.

Kanana Model's Training Efficiency and Omni Capabilities

peter, leader of Kakao's Unified Foundation Model organization, revealed the achievements of the 'Kanana' model lineup currently being developed from scratch. Notably, while global target models of similar size used 23T training tokens, Kanana achieved superior performance with just 11T tokens, half that amount, demonstrating the excellence of its data refinement. Additionally, a demonstration of the Kanana-o (Omni Model), which processes text, images, and audio in real-time, showcased its Korean language proficiency and multi-speaker conversation capabilities.

Technological Sovereignty and Hyper-Personalized Agent Strategy

Discussions emphasized the importance of operating 'high-efficiency, customized models' optimized for service environments to avoid dependency on external models. Scholars assessed that securing control over Korea's unique cultural context is essential for technological sovereignty. Kakao plans to secure market leadership through 'hyper-personalized agents' that combine 'context-aware intelligence', which predicts user behavioral context within the KakaoTalk platform, with on-device technology.

Industry-Academia Collaboration and Talent Development Support

Kakao is developing its own 'orchestration benchmark' to validate 'practical intelligence' that solves real-world problems, going beyond simple generative competition. Furthermore, to strengthen collaboration with academia, Kakao is reviewing plans to support GPU resources (in the form of credits and projects) for university professors' labs and undergraduate AI clubs.