kanana-2 embedding Models: Kakao Open-Sources 3B and 1.3B

The kanana-2 embedding release is Kakao's bet that personal AI agents need better memory search before they need bigger brains. On October 29, the Korean tech giant open-sourced two text embedding models on Hugging Face: kanana-2-3b-embedding and kanana-2-1.3b-embedding. Kakao says they beat similar-sized open models on Korean retrieval, match them on English, and run more efficiently than Qwen3-Embedding.

What actually happened

Kakao announced the release in a press note dated October 29, framing embeddings as a building block for personalized AI agents. An embedding model turns text into numbers that capture meaning, so a system can find related information even when the user does not remember the exact words.

Kakao's example is relatable: you ask an assistant for "that quiet restaurant we talked about," and it finds the right past note because the meaning matches, not the name. That is the retrieval half of retrieval-augmented generation, or RAG, and it decides whether an agent's answer is grounded in the right context.

The Hugging Face model card describes the series as Korean and English retrieval models built on Kakao's Kanana-2 small language models, converted to bidirectional attention with mean pooling. Weights for both sizes went up on October 29.

How the kanana-2 embedding models were built

Kakao says it trained the 3B model by teaching its Kanana-2 SLM to use its language understanding for search. It then distilled that knowledge into the smaller 1.3B model. The model card lists about 3.18 billion BF16 parameters for the larger model and about 1.29 billion for the smaller one.

Both use Matryoshka representation learning. In practice that lets developers shrink vectors after the fact and trade a bit of accuracy for lower storage and faster search. The card's example code shows a 4,096-token max length setting and support for sentence-transformers, transformers, and vLLM.

Agents are only as smart as what they can find. Embeddings are the finding part.

The benchmark claims

Kakao says both models scored highest on Korean among similar-sized open embedding models it tested, with English performance at a similar level and better inference efficiency than Qwen3-Embedding. The card says the 3B model has the best average on nano-BEIR (Korean) and Ko-Retrieval among compared models at every Matryoshka dimension.

Kakao also claims retrieval gains over its previous Kanana embedding series and a tens-of-times improvement in memory use and inference speed. Every evaluation ran without task prompts, per the card. These are Kakao's numbers, and independent leaderboard runs will tell whether they hold up.

For comparison points, our cohere embed 5 and pplx embed preview pieces cover the commercial side of the same race. The kanana-2 embedding drop adds an open, Korean-first option.

Why Kakao is pushing this now

This is not a one-off. Kakao open-sourced four Kanana small language models in July, and the kanana-2 embedding pair builds directly on that family. At its if(kakao)26 conference, the company already pitched embeddings that connect scattered everyday context as the starting point for personalized agents.

Noh Byung-seok, who leads Kakao's unified foundation model work, said in the release that agents need more than the ability to answer questions; they need to find the right information inside scattered data. Kakao says it plans to keep training embeddings on top of larger language models and distilling them down, aiming for better search quality at the same efficiency.

What to do as a reader

If you build search, RAG, or agent memory for Korean users, this is worth a test. Pull the 1.3B model first if latency or cost matters, then compare against the 3B on your own data. Try truncated dimensions like 256 to see how much accuracy you lose for the storage you save.

Read the Kanana Open License before shipping. Commercial use is allowed under its conditions, but check them against your product and region.

If you mostly work in English, treat Kakao's "similar English performance" as a claim to verify, not a reason to switch. Run your own eval set before replacing a production embedding model.

Sources: Kakao press release, Hugging Face kanana-2-3b-embedding, Hugging Face kanana-2-1.3b-embedding.