Kakao Unveils 'Kanana' Family of Korean-Specialized AI Models

Kakao has unveiled its AI model lineup, the 'Kanana Model Family,' aiming to respond to AGI competition and optimize services. Kanana consists of a group of models that process multimodal data including text, images, video, and audio.

The main components are as follows:

1. Language Models: Three types including Kanana Nano, Essence, and Flag. In particular, Kanana Essence is specialized for Korean language performance and shows an overwhelming advantage over global models.
2. Multimodal Language Models: Composed of Kanana-v (vision), Kanana-a (audio), Kanana-o (omni), etc., and is being developed with an End-to-End structure. The currently completed TAI2TA structure generates audio and text responses within an average of 1.6 seconds.
3. Visual Generation Models: Includes 'Kollage by Kanana' for high-quality image generation and 'Kinema by Kanana' for video content generation.
4. Speech Models: Composed of the real-time streaming speech recognition model 'Karve by Kanana' and the zero-shot speech synthesis model 'Kast by Kanana'.

The Kanana models are trained on data where copyright and personal information issues have been resolved to ensure transparency, and achieve excellent performance even at small sizes through service-specific tuning such as summarization, RAG, and Function Calling. They are currently applied to KakaoTalk AI Summary, Talk Channel AI Manager, etc., and are planned to expand to the Kanana app, Daum, Shopping, etc., in the future.