Kakao Unveils Evolution of Agentic AI-Oriented 'Kanana-2' and Omni Model 'Kanana-o'

Kakao unveiled the development status and achievements of the language model 'Kanana-2' and the multimodal omni model 'Kanana-o', aiming for evolution into Agentic AI, at if(kakao)25.

Kanana-2: Balancing Performance and Efficiency

The core language model Kanana-2 applies MLA (Multi-head Attention) and MoE (Mixture of Experts) structures to maximize efficiency in long-input processing and large-scale inference environments. The currently under-development Kanana-2-30b-a3b-preview model has recorded groundbreaking inference throughput compared to previous models, enabling faster responses with fewer resources in real-service environments. Additionally, an 8B lightweight model is being developed for on-device application, with plans to expand to various sizes in the future through Pruning and Distillation.

Kanana-o: Korean-Specialized Any-to-Any Model

Kanana-o, which integrates understanding of text, voice, and images, was developed by combining the visual understanding model Kanana-v and the voice generation model Kanana-a. Trained on Korean and culture-specific datasets, it has demonstrated outstanding performance on Korean benchmarks and shows competitive performance against global top-tier models such as GPT-4o. By applying streaming technology, it enables real-time conversation without latency, and is scheduled to be further advanced with Full-duplex conversation support in the future.

Multimodal Generation and Ecosystem Expansion

The image generation model Kanana-kollage has already been applied to apps such as Kanana App, and the video generation model Kanana-kinema, based on it, supports natural video production. The voice language model is scheduled to expand to support a total of 5 languages (Korean, English, Japanese, Chinese, Vietnamese). Based on this technical foundation, Kakao plans to fully apply Kanana models to its own services, such as KakaoTalk, to spread the Agentic AI experience.