Kakao 在 Hugging Face 开源两款韩语优化 AI 搜索嵌入模型
Kakao Open-Sources Two AI Search Models Optimized for Korean
Kakao 于 10 月 29 日在 Hugging Face 开源 Kanana-2-3b-embedding（30 亿参数）和 Kanana-2-1.3b-embedding（13 亿参数）两款嵌入模型，面向个性化 AI 智能体与 RAG 场景。Kakao 称内部评测中两款模型韩语检索性能优于同规模 llama-nemotron-embed-1b-v2 和 Qwen3-Embedding-4B，英语性能相当，推理效率更高，内存占用与推理速度较前代 Kanana 嵌入模型提升数十倍以上。模型采用 MRL 支持开发者调节嵌入向量大小，并以 Kanana Open License 允许商用。
正文
Kakao released two open-source embedding models on Oct. 29 for use in developing personalized AI agents. (Image: Kakao)
PC본문_6월 Mouser (in Article)//
AD Manager | AD178089411676 //
// AD Manager | AD178089411676
Kakao has released two open-source embedding models designed for developing personalized artificial intelligence (AI) agents. The models feature improved Korean-language search performance and smaller model sizes to enhance operational efficiency.
Kakao said on Oct. 29 that it had released two Kanana-2-embedding models on Hugging Face, a global AI model-sharing platform. The release expands access to its proprietary AI technology following the company's publication of four small language models (SLMs) in July.
The newly released models are Kanana-2-3b-embedding, with 3 billion parameters, and Kanana-2-1.3b-embedding, with 1.3 billion parameters.
Embedding is a technology that converts the meaning of sentences into numerical representations to retrieve relevant information. It can identify semantically similar information even when exact words do not match, making it useful for personalized AI agents and retrieval-augmented generation (RAG) systems.
For example, even if a user cannot remember the name of a restaurant mentioned in a previous conversation, the system can retrieve relevant information from a query such as "the quiet restaurant we talked about before." Kakao plans to use the technology as a core foundation for personalized AI agents that connect users' past conversations with relevant information.
According to Kakao, internal evaluations showed that both models delivered stronger Korean-language search performance than similarly sized global open models, including llama-nemotron-embed-1b-v2 and Qwen3-Embedding-4B. English-language performance was comparable, while inference efficiency was also higher than that of Qwen3-Embedding.
Search performance also improved compared with Kakao's previous Kanana embedding models. The company said it confirmed efficiency gains of several dozen times or more in memory usage and inference speed.
Kakao used its proprietary small language model, Kanana-2-SLM, to develop the new models. It first developed the 3 billion-parameter model by applying language understanding capabilities to retrieval tasks, then built the 1.3 billion-parameter model using knowledge distillation, a technique that transfers learning from a larger model to a smaller one.
The company also incorporated technology that allows developers to balance search performance and resource consumption according to their service environments. Using Matryoshka Representation Learning (MRL), developers can adjust the size of embedding vectors to optimize storage requirements and processing speed.
Both models are released under the Kanana Open License, which permits commercial use subject to its licensing terms. Companies and developers in South Korea and overseas can use the models to build semantic search services and personalized AI agents.
Kakao plans to expand research into embedding training and knowledge distillation using larger language models.