Kakao Open-Sources 2 Kanana-2 Embedding Models

Kakao has released two open embedding models, Kanana-2-3b-embedding and Kanana-2-1.3b-embedding, built to find information by meaning in Korean and English. The weights are on Hugging Face under Kakao’s own licence, which allows commercial use.

The South Korean internet company announced the release in a Korean-language statement and called embedding the base technology for its personalised AI agents. It follows four small Kanana language models that Kakao open-sourced in July.

What the models do

An embedding model turns text into a list of numbers that captures its meaning, so a system can find related information even when the user does not remember the exact words. Kakao’s example is a user asking for “the quiet restaurant we talked about before”. The model can match that to an earlier note about “a good place for a conversation”, without the restaurant’s name.

That kind of search underpins retrieval-augmented generation, where an AI model looks up relevant information before answering. Kakao says it wants its agents to draw on the scattered context of a user’s daily life, an idea it set out at its if(kakao)26 conference.

The larger model has 3bn parameters and the smaller one 1.3bn. Kakao built the 3bn model by training its Kanana-2 small language model for retrieval, then made the 1.3bn model by distillation, a method in which a large model teaches a smaller one. Both accept up to 4,096 tokens of input.

The models also use Matryoshka representation learning. This lets developers cut the output from its full size, 2,560 numbers for the larger model and 1,280 for the smaller, down to 1,024, 512, 256 or 128. Shorter vectors use less storage and search faster, at some cost to accuracy. Cut to 128, the 3bn model’s Ko-Retrieval score falls from 0.8211 to 0.7488, still above the 0.6854 that Qwen3-Embedding-4B scores at the same size.

The benchmark results

According to the model card, Kakao tested the models with the MTEB evaluation tools and no task prompts. At full size, the 3bn model scored:

0.8211 on Ko-Retrieval, an average of six Korean search tasks, against 0.7837 for Alibaba’s Qwen3-Embedding-4B and 0.7985 for NVIDIA’s llama-nemotron-embed-1b-v2;
0.6088 on the Korean version of nano-BEIR, against 0.5522 for Qwen3-Embedding-4B;
0.6317 on MTEB’s English benchmark, behind Qwen3-Embedding-4B’s 0.6661.

The 1.3bn model scored 0.7958 on Ko-Retrieval. Kakao also says the new models are tens of times more efficient in memory use and speed than its earlier embedding models, but it did not publish those figures.

“For AI agents to give users the help they need, they need not only the ability to answer questions but also technology that accurately finds the necessary information among scattered data,” said Noh Byung-seok, who leads Kakao’s unified foundation model work, in a statement translated by TSN. He said the release would help researchers and developers use embedding technology that is strong in Korean across a range of services.

Korean search is the selling point

Kakao’s own numbers make the trade-off clear. The models lead on Korean retrieval, the job Kakao built them for, while Qwen3-Embedding-4B remains ahead on English. Kakao’s release calls the English results “similar”. For a Korean company or developer building search over Korean documents, a 3bn model that beats a 4bn rival on home-language tasks is a practical option, The smaller 1.3bn model is a closer call: it beats NVIDIA’s 1bn model on Korean nano-BEIR, 0.5866 to 0.5471, but trails it slightly on Ko-Retrieval, 0.7958 to 0.7985.

All the scores are Kakao’s own, run without prompts, and independent tests may differ. The release also adds to a wave of small open-weight models aimed at cheaper deployment.

For Kakao, open-sourcing embedding models builds developer interest in its stack as Korean rivals push their own AI plans, from SK Telecom’s agent workforce to Naver’s $9bn compute deal with Brookfield. Kakao says it will next train embedding models on larger language models.