[Edaily Reporter Kim Hyun-ah ] Kakao(035720)has unveiled two artificial intelligence (AI) models that excel in Korean-language search performance. These models scored higher than comparable models in Korean search evaluations and are designed to reduce the size of search data, thereby minimizing storage space requirements and processing burdens. Companies can also use them commercially under the terms of the license.

Kakao announced that it has released two versions of its “Kanana-2-embedding” model on the open-source platform Hugging Face. The released models have 3 billion (3B) and 1.3 billion (1.3B) parameters, respectively.

Embedding models are a technology that represents the meaning of sentences or documents as numerical values to help find relevant information. Even if a search term and a document do not contain the exact same words, the model can identify relevant information if their meanings are similar. This technology serves as the foundation for internal corporate document search systems and AI agents designed to understand user-specific contexts.

Kakao highlights its Korean search performance as a key differentiator. According to the company’s evaluation, the 3B model achieved the highest average score among the compared models in the Korean search category of the Nano-BEIR Korean evaluation.

The comparison models included NVIDIA’s “Llama-nemotron-embed-1b-v2” and Alibaba’s “Qwen3-Embedding-4B,” among others. Kakao explained that its English performance was also on par with that of the comparison models.

Technology to reduce search costs was also implemented. By introducing the Matryoshka Representation Learning (MRL) technique, the model allows the dimensions of embedding vectors to be adjusted to 1,024, 512, 256, or 128. As the vector dimension decreases, the storage space required for search information and the processing load are reduced. Developers can select the vector size based on the search performance and cost requirements of their service.

The open-source model is licensed under Kakao’s proprietary “Kanana Open License.” Commercial use is permitted provided the license terms are met. Companies and developers can utilize this model to develop search systems or AI agents.

Noh Byung-seok, Performance Leader for Kakao’s Unified Foundation Model, stated, “For AI agents to provide users with the help they need, it is important not only to answer questions but also to accurately identify the necessary information from scattered data.” He added, “We will support the application of embedding technology—which excels in Korean—across a variety of services.”