Kakao on Jan. 5 unveiled the performance of its in-house AI model "Kanana-v-4b-hybrid," which can handle everything from everyday conversations to complex problems that require logical reasoning. The model was developed based on "Kanana-1.5-v-3b," which was released as open source last year, and unlike the previous model, it has the ability to synthesize information and compute like a person. Through a self-check process, it significantly improved accuracy by reducing calculation errors and missing conditions.
"Kanana-v-4b-hybrid" improved its performance through a four-stage, advanced training procedure that includes foundational training, long-form chain-of-thought, and offline and online Reinforcement Learning. In particular, it achieved outstanding results in Korean logical development capability and recorded a high correct-answer rate on College Scholastic Ability Test questions. The model scored 92.8 on KoNET (Korea National Educational Test Benchmark), demonstrating its ability to accurately understand and reason in response to Korean-language questions.
Kakao said the model will be advanced into a form that can automatically switch between general conversation mode and reasoning mode, providing users with a natural and efficient user experience. Kakao also noted it will continue to develop models specialized in Korean to gain an edge in global AI competition.