카카오 ‘카나나2' 모델 오픈소스 공개[사진=카카오]

<이미지를 클릭하시면 크게 보실 수 있습니다>

SEOUL, December 19 (AJP) - South Korea's Kakao Corp. said on Friday it has released its next-generation large language model, Kanana-2, as open source.

The company said it published the in-house model on Hugging Face, adding that the release showcases technology optimized for building AI agents capable of understanding context and executing tasks autonomously.

Kakao began unveiling its Kanana model lineup last year and has since released successive versions as open source, ranging from lightweight models to Kanana 1.5, which was designed to address more complex problem-solving tasks.

Kanana-2 incorporates Kakao’s latest research aimed at significantly improving performance and efficiency, with a focus on so-called “companion” AI that can interpret user intent and act proactively.

According to Kakao, Kanana-2 delivers substantial improvements in tool-calling and instruction-following, two core capabilities for agentic AI systems. Compared with its predecessor, the model achieved more than a threefold improvement in multi-turn conversational tool-calling performance and was designed to execute complex, step-by-step tasks. Language support was expanded to six languages, including Korean.

The company said it adopted newer architectures to boost efficiency, including multi-latent attention to process long input sequences and a mixture-of-experts structure that activates only selected parameters during inference.

“Innovative AI services are built on the performance and efficiency of language models,” said Kim Byung-hak, Kakao’s Kanana performance lead. “Beyond high performance, we will develop practical AI models that operate quickly and effectively in real service environments, and continue sharing them as open source to help energize AI research ecosystems at home and abroad.”
Baek Seo-hyun 기자 qortjgus0602@ajunews.com

- Copyright ⓒ [아주경제 ajunews.com] 무단전재 배포금지 -