South Korean Tech Firm Kakao Releases Open-Source Kanana-2 Embedding Models for Commercial AI Agents

South Korean technology company Kakao released two open-source embedding models named “Kanana-2-embedding” on the Hugging Face platform on October 29, designed to provide foundational support for personalized artificial intelligence agents (AI Agents). The models include “Kanana-2-3b-embedding” with 3 billion parameters and “Kanana-2-1.3b-embedding” with 1.3 billion parameters. Embedding models convert text semantics into numerical vectors, enabling AI systems to retrieve relevant information based on meaning similarity even when users do not input precise keywords, which is a key component for building AI agents with long-term memory and context understanding.

According to Kakao, the new models utilize its proprietary language models to improve search accuracy and operational efficiency. In internal comparative evaluations against existing Kanana series and similar public models (such as llama-nemotron-embed-1b-v2 and Qwen3-Embedding-4B), Kanana-2 demonstrated the best performance in Korean language tasks, comparable performance in English, and higher inference efficiency than Qwen3-Embedding. Additionally, the models employ MRL (Matryoshka Representation Learning) technology, allowing developers to dynamically adjust vector dimensions based on storage and processing needs to balance performance and resource consumption.

The released models follow the “Kanana Open License,” permitting commercial use under specific conditions. This continues Kakao’s strategy from July this year when it open-sourced lightweight language models, further enriching the Korean AI infrastructure ecosystem. No Byung-seok, Performance Leader at Kakao Unified Foundation Model, stated that the initiative aims to help researchers and developers apply strong Korean-capable embedding technologies to various services while continuously optimizing search performance and efficiency.