Kakao Open-Sources AI Agent-Specialized LLM 'Kanana-2'

Attention, AI developers! Kakao has officially open-sourced its self-developed next-generation language model, 'Kanana-2', on Hugging Face. This model, designed for high performance and high efficiency, is optimized for implementing Agentic AI and is expected to open new possibilities in AI service development. 'Kanana-2' consists of three types: the 'Base' model, the 'Instruct' model which enhances instruction-following capabilities, and the 'Thinking' specialized model, making its debut this time. Notably, all weights from the training phase have been released, allowing developers to fine-tune the model with their own data, significantly expanding the possibilities for custom AI development. The biggest innovation of 'Kanana-2' is the significant enhancement of its 'Tool Calling' function, a core aspect of Agentic AI, and its ability to execute user instructions. Its multi-turn conversation tool-calling capability has improved by more than three times compared to previous models, and the supported languages have expanded to a total of six, including Japanese, Chinese, Thai, and Vietnamese, in addition to Korean and English, increasing global usability. Technologically, it adopts the 'MLA (Multi-head Latent Attention)' technique for efficiently processing long inputs and the 'MoE (Mixture of Experts)' architecture, which activates only the necessary parameters during inference. This allows it to quickly process long contexts with fewer memory resources, achieving both reduced computational costs and improved response speed. Its performance is also top-tier. The 'Instruct' model stands shoulder to shoulder with the latest models of the same architecture, and the 'Thinking' specialized model has also demonstrated excellent performance in benchmarks requiring diverse reasoning abilities. Its practical utility has already been verified in real development environments through the 'AI Agent Competition'. Kakao has consistently led the open-sourcing of AI models through its 'Kanana' series. Moving forward, Kakao announced its ambition to contribute to the revitalization of the domestic and international AI research ecosystem by continuously pursuing the expansion of MoE-architecture-based model scales, development of AI agent scenario-specific models, and advancement of on-device lightweight models. Kim Byung-hak, Performance Leader of Kakao Kanana, emphasized, "We will consistently share practical AI models that can be applied to real AI services." 'Kanana-2' is more than just a language model; it will serve as a powerful foundation for developers to implement innovative Agentic AI services. Discover Kakao's technology on Hugging Face right now!