Kakao开源Kanana-2小语言模型，主打韩英双语与端侧运行

Kakao has released the small language models (SLMs) of its in-house artificial intelligence (AI) model "Kanana" as open source, marking its full-fledged entry into building an AI ecosystem.
Kakao (035720) said on the 28th that it had open-sourced four SLMs on the global AI platform Hugging Face that can run on devices such as smartphones. The models released are the "Kanana-2" small language model series, comprising four in total: Kanana-2-1.3B-base, Kanana-2-1.3B-instruct, Kanana-2-3B-base, and Kanana-2-3B-instruct.
In the AI industry, the importance of small language models that can run directly on devices such as smartphones or personal PCs has recently come into focus. Global companies are seeking to gain an early lead in the lightweight AI ecosystem, promoting on-device AI as a key strategy or successively releasing lightweight open-source models. Amid this trend, Kakao has continuously researched small language models with performance levels applicable to actual services.
The models released this time are notable for delivering global-level performance in both Korean and English despite their small size. Compared with state-of-the-art (SOTA) open-source models of similar size, they recorded superior performance on most major benchmarks, including Korean, English, knowledge, math, and code. In particular, the open-sourced Kanana-2-1.3B-instruct, along with the smaller Kanana-2-0.9B-instruct developed alongside it, showed competitive performance against global models of the same class, such as Qwen and Gemma, in practical service capabilities including conversation, knowledge, code, math, instruction-following, and tool calling. In addition, Kakao (035720) applied its in-house Korean-specialized tokenizer to improve Korean processing efficiency by more than 30% compared with before.
Another advantage is that the models are optimized for on-device environments with limited memory and computing resources. To solve the problem of memory usage surging when processing long conversations, the models adopted a "sliding window attention" structure, which allows them to maintain stable performance while reducing memory use by up to 72.7% even in conversations of up to 32K (32,000 tokens, about 24,000 words) in length.
Kakao is distributing these 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 company expects to contribute to expanding the foundation for using high-performance Korean AI models domestically and to invigorating the development of new services.
Noh Byung-seok, performance leader of Kakao's Unified Foundation Model, said, "As we prepare for the era of agentic AI, we are keenly aware of the importance of both large-scale AI in the cloud and lightweight on-device AI." He added, "We hope these models will lead more developers and companies to develop new AI services, contributing to invigorating the domestic AI ecosystem."