Kakao Open-Sources Two Kanana Embedding Models for Personalized AI Agents

Kakao said on the 29th that it had open-sourced two versions of "Kanana-2-Embedding," an embedding model that serves as foundational technology for personalized AI agents.

The newly released models are "Kanana-2-3b-Embedding," with 3 billion parameters, and "Kanana-2-1.3b-Embedding," with 1.3 billion parameters.

Kakao said on the 29th that it had open-sourced two versions of “Kanana-2-Embedding,” an embedding model that serves as foundational technology for personalized AI agents. Kakao

An embedding model represents the meaning of text numerically and uses it to find information relevant to a user's question. While a general language model understands questions and generates answers, an embedding model finds the information needed to answer them based on the semantic similarity between questions and information. For example, when a user looks for "that quiet restaurant we talked about before," the model can help find past information with a similar meaning, such as "a good place to talk," even if the user does not enter the restaurant's name.

In evaluations against "llama-nemotron-embed-1b-v2" and "Qwen3-Embedding-4B," publicly available embedding models of similar sizes, the two released models delivered the highest performance in Korean and English performance at a comparable level. Compared with Kakao's previous Kanana embedding model series, they also showed efficiency gains of dozens of times or more in terms of memory use and inference speed.

Kakao developed the 3B model by training its proprietary small language model, "Kanana-2-SLM," to apply its language-understanding capabilities to information retrieval. It then built the 1.3B model using knowledge distillation, a technique that transfers knowledge learned by a larger model to a smaller one.

Both models are released under the Kanana Open License, which allows commercial use subject to the license terms. Researchers and companies in Korea and abroad can directly evaluate the publicly released models and use them to develop personalized AI agents and semantic search services.

Noh Byeongseok, Unified Foundation Model Performance Lead at Kakao, said, "For AI agents to provide users with the help they need, it is important not only to answer questions but also to accurately find the necessary information among scattered data. Through this release, we will help researchers and developers apply embedding technology with strengths in Korean to a variety of services, and continue research to improve both search performance and efficiency."

This content was produced with the assistance of AI translation services.