Performance comparison of four lightweight language models Kakao released as open source on the 28th. Photo courtesy of Kakao
Kakao has open-sourced the small language models (SLM) of its in-house developed artificial intelligence (AI) model “Kanana.”
Kakao (035720) said on the 28th that it has open-sourced four Kanana-2 series SLMs, which can run on device environments such as smartphones, on the global AI platform Hugging Face.
In the AI industry recently, the importance of lightweight language models that can run directly in environments such as smartphones and personal PCs has been highlighted. Global companies are focusing on gaining an early lead in the lightweight AI ecosystem, either by presenting on-device AI as a key strategy or by successively releasing lightweight open-source models. Amid this trend, Kakao has continuously researched lightweight language models with performance levels applicable to actual services.
The models released this time are characterized by achieving global-level performance in both Korean and English despite their small size. Compared to the latest state-of-the-art (SOTA) open-source models of similar size, they recorded superior performance in most representative benchmarks, including Korean, English, knowledge, math and code. In particular, the open-sourced “Kanana-2-1.3B-Instruct” and the smaller “Kanana-2-0.9B-Instruct” showed competitive performance against global models in the same class, such as Qwen and Gemma, in actual service capabilities including conversation, knowledge, code, math, instruction-following and tool calling. In addition, Kakao (035720) applied its in-house developed Korean-specialized tokenizer to improve Korean processing efficiency by more than 30% compared to before.
It is also notable that the models are optimized for on-device environments with limited memory and computing resources. To solve the problem of memory usage surging sharply when processing long conversations, a “sliding window attention” structure was introduced, allowing memory to be reduced by up to 72.7% while maintaining stable performance even in conversations of up to 32K (32,000 tokens, about 24,000 words) in length.
Kakao is distributing the models under the “Kanana Open License,” which permits even commercial use. Anyone, including developers, startups and research institutions, can use the technology without separate restrictions. Through this, the models are expected to contribute to expanding the foundation for using high-performance Korean AI models domestically and to invigorating the development of new services.
Noh Byoung-seok, performance leader of Kakao’s Unified Foundation Model team, said, “As we prepare for the era of agentic AI, we are feeling the importance of both large-scale cloud AI and lightweight on-device AI.” He added, “We hope these models will lead to the development of new AI services by more developers and companies, contributing to the revitalization of the domestic AI ecosystem.”