AI · 2026-11-01
Kakao unveils two Korean-specialized Kanana embedding models… even past conversations can be searched by meaning
Kakao has released two Kanana embedding models as open source, optimized for Korean and English search. The technology helps find information with similar meaning even when the exact words do not match, and it can be used in personalized AI agents and enterprise document search.

If you asked, “Find that quiet restaurant we talked about before,” based on a past conversation, could the system locate the right information even if you no longer remember the restaurant’s name or the exact wording from that time? One technology that makes this kind of search possible is embedding. By releasing two Kanana embedding models specialized for Korean and English search, Kakao has expanded the use of its in-house language models into personalized AI and information retrieval.
The models Kakao published on Hugging Face are the 3-billion-parameter “Kanana-2-3B-Embedding” and the 1.3-billion-parameter “Kanana-2-1.3B-Embedding.” Parameters are the elements a model uses to learn and process information, and they are often used to gauge the scale of a model. After previously releasing its small language model, Kakao is now introducing embedding models for search and information discovery.
Embedding is a technique that converts the meaning of a sentence or document into a numerical vector. Keyword search focuses on finding information that matches the words entered in the query, while semantic search can find relevant information even when the wording is different, as long as the meaning is similar. For example, if a restaurant was recorded in a past conversation as a “good place to talk” and a user later searches for a “quiet restaurant,” the system can retrieve that information based on the semantic relationship between the two expressions.
That said, an embedding model itself does not remember all of a user’s conversations or generate answers on its own. Its role is to find information in stored data that is highly relevant to the current question and pass it on to the language model. This approach, in which retrieved documents or information are used to generate an answer, is called retrieval-augmented generation, or RAG. Search results matter in enterprise document search, customer support, and personalized AI services because they can affect the accuracy and relevance of the final response.
Kakao said it trained the models to use the language understanding capabilities of its own Kanana-2 small language model for information retrieval. In addition to semantic search, the two models can also be used for sentence similarity analysis and RAG. The 1.3-billion-parameter model uses knowledge distillation, a technique that transfers knowledge from a larger model to a smaller one, allowing developers to choose a model based on service compute resources and cost considerations.
Another notable feature is that the vector size can be adjusted. According to the published details, the default embedding size is 2,560 dimensions for the 3-billion-parameter model and 1,280 dimensions for the 1.3-billion-parameter model, and the models also offer the option of using smaller dimensions through Matryoshka representation learning. Lowering the vector dimension can reduce storage requirements and search computation overhead, although actual search quality may vary depending on the configuration and data.
In the evaluation materials released by Kakao, both models scored higher than the comparison models on Korean search tasks. The 3-billion-parameter model scored 0.6088 on the Nano-BEIR Korean evaluation and 0.8211 on the CoReTRieval evaluation, which combines six Korean search tasks. The 1.3-billion-parameter model also scored 0.5866 on the Nano-BEIR Korean evaluation. English performance, however, varied by benchmark. The 3-billion-parameter model scored 0.6552 on Nano-BEIR English, but on MTEB v2 English it scored 0.6317, lower than the 0.6661 of Qwen3-Embedding 4B, which was used for comparison. These figures are based on benchmark data released by Kakao, and actual service performance may differ depending on the data and system setup.
Kakao said it applied the Kanana open license, which allows commercial use under the stated licensing terms in addition to research and development use. However, developers and companies planning commercial use will need to review the specific license conditions.
This release shows that in the AI race, it is becoming increasingly important not only to generate natural answers, but also to retrieve the right context from vast amounts of conversation and documents with precision. The Kanana embedding models widen the options available for building personalized AI agents and enterprise search systems. Their real-world competitiveness will likely be judged not only by benchmark scores, but also by how well search quality and operational efficiency are implemented in actual service environments.
The models Kakao published on Hugging Face are the 3-billion-parameter “Kanana-2-3B-Embedding” and the 1.3-billion-parameter “Kanana-2-1.3B-Embedding.” Parameters are the elements a model uses to learn and process information, and they are often used to gauge the scale of a model. After previously releasing its small language model, Kakao is now introducing embedding models for search and information discovery.
Embedding is a technique that converts the meaning of a sentence or document into a numerical vector. Keyword search focuses on finding information that matches the words entered in the query, while semantic search can find relevant information even when the wording is different, as long as the meaning is similar. For example, if a restaurant was recorded in a past conversation as a “good place to talk” and a user later searches for a “quiet restaurant,” the system can retrieve that information based on the semantic relationship between the two expressions.
That said, an embedding model itself does not remember all of a user’s conversations or generate answers on its own. Its role is to find information in stored data that is highly relevant to the current question and pass it on to the language model. This approach, in which retrieved documents or information are used to generate an answer, is called retrieval-augmented generation, or RAG. Search results matter in enterprise document search, customer support, and personalized AI services because they can affect the accuracy and relevance of the final response.
Kakao said it trained the models to use the language understanding capabilities of its own Kanana-2 small language model for information retrieval. In addition to semantic search, the two models can also be used for sentence similarity analysis and RAG. The 1.3-billion-parameter model uses knowledge distillation, a technique that transfers knowledge from a larger model to a smaller one, allowing developers to choose a model based on service compute resources and cost considerations.
Another notable feature is that the vector size can be adjusted. According to the published details, the default embedding size is 2,560 dimensions for the 3-billion-parameter model and 1,280 dimensions for the 1.3-billion-parameter model, and the models also offer the option of using smaller dimensions through Matryoshka representation learning. Lowering the vector dimension can reduce storage requirements and search computation overhead, although actual search quality may vary depending on the configuration and data.
In the evaluation materials released by Kakao, both models scored higher than the comparison models on Korean search tasks. The 3-billion-parameter model scored 0.6088 on the Nano-BEIR Korean evaluation and 0.8211 on the CoReTRieval evaluation, which combines six Korean search tasks. The 1.3-billion-parameter model also scored 0.5866 on the Nano-BEIR Korean evaluation. English performance, however, varied by benchmark. The 3-billion-parameter model scored 0.6552 on Nano-BEIR English, but on MTEB v2 English it scored 0.6317, lower than the 0.6661 of Qwen3-Embedding 4B, which was used for comparison. These figures are based on benchmark data released by Kakao, and actual service performance may differ depending on the data and system setup.
Kakao said it applied the Kanana open license, which allows commercial use under the stated licensing terms in addition to research and development use. However, developers and companies planning commercial use will need to review the specific license conditions.
This release shows that in the AI race, it is becoming increasingly important not only to generate natural answers, but also to retrieve the right context from vast amounts of conversation and documents with precision. The Kanana embedding models widen the options available for building personalized AI agents and enterprise search systems. Their real-world competitiveness will likely be judged not only by benchmark scores, but also by how well search quality and operational efficiency are implemented in actual service environments.
Related search terms
#Kakao Kanana embedding model#Korean embedding model#semantic search model#AI search for conversations#RAG for enterprise search#Hugging Face Korean AI model#personalized AI agent search