On the 27th, Kakao released a technical report detailing the research achievements of its self-developed language model 'Kanana' and distributed the lightweight language model 'Kanana NANO 2.1B' as open source on GitHub. This report includes detailed information on the structure of the Kanana model, learning strategies, and performance in global benchmarks.
Kakao completed the training of the super-large language model 'Kanana Flag' at the end of last year and established a full lineup of Kanana language models. The Kanana Flag outperformed other models in Korean language performance and recorded similar results to competing models in English performance. Kakao plans to enhance reasoning, mathematical, and coding capabilities by applying reinforcement learning and continuous learning technologies in the future and to improve the accuracy of user request fulfillment.
On this day, Kakao released the base model, instruct model, and embedding model of the 'Kanana NANO 2.1B' to the open-source community. This model is a high-performance lightweight model that can be utilized by researchers and developers, even in on-device environments. Kakao expects that this open-source release will contribute to activating the domestic AI ecosystem.