Kakao has unveiled its next-generation language model optimized for agentic AI, enhancing its proprietary AI technology capabilities. On the 19th, Kakao announced through Hugging Face the open-source release of its latest language model, ‘Kanana-2.’ With significantly improved performance and efficiency, this model showcases Kakao’s agentic AI technology, enabling it to actively understand and execute user instructions.
Since introducing its AI model lineup ‘Kanana’ last year, Kakao has sequentially open-sourced various models, ranging from lightweight models to ‘Kanana-1.5,’ designed for solving complex problems. The newly introduced ‘Kanana-2’ incorporates extensive research outcomes, focusing on performance enhancement along with efficiency and practicality for real-world service applications.
The ‘Kanana-2’ model comprises three versions: a base model with fundamental performance, an instruct model enhanced for executing user commands through post-training, and for the first time, a thinking model specialized in reasoning. By fully disclosing the training weights, developers can freely fine-tune the model with proprietary data, increasing its applicability in both research and service sectors.
‘Kanana-2’ significantly strengthens key aspects of agentic AI, such as tool calling and instruction execution capabilities. It boosts multi-turn tool calling performance over three times compared to the previous model ‘Kanana-1.5-32.5b’ and is designed to accurately understand and perform complex step-by-step requirements. The supported languages have expanded from Korean and English to include Japanese, Chinese, Thai, and Vietnamese.
Technically, the model maximizes efficiency by implementing the latest architecture, incorporating ‘MLA (Multi-head Latent Attention)’ for efficient long input processing, and ‘MoE (Mixture of Experts)’ structure that activates only necessary parameters during inference. This reduces memory usage and improves computational cost and response speed, achieving high throughput even in large-scale concurrent request environments.
In terms of performance, the instruct model matches the performance of the latest similar-structured model, ‘Qwen3-30B-A3B,’ while the thinking-specialized model demonstrates similar reasoning capabilities against key benchmarks requiring cognitive abilities. These models were pre-launched for participants at the ‘AI Agent Competition’ co-hosted with the Korean Institute of Information Scientists and Engineers this month, showcasing their potential utility in actual agent development environments.
The release of ‘Kanana-2’ symbolizes Kakao’s strategy to bridge research outcomes with competitive service capabilities. Under the leadership of Byeonghak Kim, who oversees the ‘Kanana’ project, Kakao plans to further develop models with advanced instruction-following ability using the MoE structure. They also aim to enhance models specialized in complex AI agent scenarios and on-device lightweight models.
‘Kanana’ stands as Kakao’s integrated AI service brand, embodying the concept of being the “most true-to-self AI.” It aspires to be an ‘AI Mate’ that understands context in everyday conversation environments like KakaoTalk, offering functionality such as conversation summarization, schedule and task management, Q&A, search, and voice recognition. Kakao continues to advance personalized AI experiences through ‘Kanana.’