Kakao will open to the public its own small language model (SLM) that can run on smartphones and personal computers. The strategy is to secure technological competitiveness not only in large language models (LLMs) but also in the On-device AI field, and to expand the foundation for use in Korea's development ecosystem.
Kakao on the 28th released four lightweight Kanana 2 models via the global AI platform Hugging Face. They consist of base and instruct models with 1.3 billion and 3 billion parameters. All are covered by the Kanana Open License, which allows commercial use, so corporations, research institutions, and developers can use them to build services.
Kakao equipped the new models with a Korean-specialized tokenizer it developed in-house. It said it improved the way sentences are split into units that AI can process to suit Korean, increasing processing efficiency by more than 30% compared with before. Because the same sentence can be converted into fewer tokens, it reduces computation time and expense.
It also applied a sliding window attention technology that uses limited memory efficiently. While processing conversations up to 32,000 tokens long, it can cut memory usage by up to 72.7%. The company said it secured performance competitive with similar-sized models such as Qwen and Gemma in key evaluations including Korean and English, math, coding, and tool use.
Kakao is using the lightweight AI models for KakaoTalk chat and call summaries, Kanana in KakaoTalk, and the AI National Secretary. Noh Byung-seok, performance lead for Kakao's Unified Foundation Model, said, "We will strengthen both cloud-based large AI and On-device AI capabilities to help activate domestic AI service development."