Kakao open-sources two Kanana-2 embedding models, built for Korean search

What this is about

Kakao, one of South Korea's big internet companies, released two embedding models on Hugging Face on October 29, 2026. They are called Kanana-2-3b-embedding (3 billion parameters) and Kanana-2-1.3b-embedding (1.3 billion parameters). Parameters are, roughly speaking, the dials a model adjusts during training. According to Kakao, the model architecture, source code and key weights are openly available. Anyone can inspect, study, modify and redistribute them. The source doesn't name the exact license.

Kakao positions the models as a foundational technology for personalized AI agents. In July, the company had already open-sourced four small language models (SLMs) in the Kanana family.

What an embedding model does

An embedding model turns the meaning of a text into a sequence of numbers, called a vector. A regular language model understands a question and answers it. An embedding model instead compares the meaning of the question with existing texts and finds relevant information that way. Texts with similar meanings have similar vectors.

That makes search work even when you don't remember the exact words. What counts is that the meaning is close. Models like this often serve as the search component in systems that feed knowledge to an AI, such as document search or agents that access personal data. These use cases are general context, not part of Kakao's announcement.

Performance and efficiency

Kakao compared the models with similarly sized open embedding models, namely llama-nemotron-embed-1b-v2 and Qwen3-Embedding-4B. Kakao's result: the best performance in Korean and similar performance in English. Compared with Qwen3-Embedding, they also showed higher inference efficiency, meaning they run faster or leaner in use. The source gives no concrete benchmark numbers. All figures come from the vendor and haven't been independently verified.

Search performance is said to be better than in the previous Kanana embedding series. In memory use and speed, Kakao reports an improvement of many times over (more than tenfold), as the source puts it.

Flexible vector size with MRL

The models use Matryoshka Representation Learning (MRL). The name refers to Russian matryoshka nesting dolls. The model is trained so that the first numbers in the vector already carry the most important information. You can therefore truncate the vector and lose only a little quality. Developers can pick the length they need and balance search quality, storage and processing speed. Shorter vectors save memory and speed up search in large databases.

What you need to try it

The source doesn't list hardware requirements. As a rough estimate, a 1.3-billion-parameter model runs in 16-bit precision with a few GB of VRAM (video memory). A quantized version also runs on a regular CPU or a Mac. The 3-billion-parameter model needs about 6–8 GB in 16-bit. Also plan for a few GB of disk space. Google Colab is often enough for tests. These are estimates, not vendor figures.

Our take

Embedding models with strong Korean support are even rarer among open models than English ones. That makes the Kanana models interesting for developers working in Korean. The MRL option makes them practical for production systems with memory limits. Some caveats remain: the comparisons come from Kakao itself, the announcement gives no numbers or test datasets, and the license is unclear. How well the models handle German is unknown. Kakao plans to continue its research with larger language models as a base and with knowledge distillation, which means transferring knowledge from a large model to a small one.