MetaPatchAI home
AI

Kakao Open-Sources Kanana-2 Embedding Models, and They Top Its Korean Search Benchmarks

Compiled by MetaPatchAI · Published , 3:25 PM ET

Oct 30, 2026: South Korea's Kakao has released two open embedding models built for one job: finding the right information by meaning, not exact keywords. On Oct 29, 2026 (KST), Kakao published Kanana-2-3b-embedding and Kanana-2-1.3b-embedding on Hugging Face under its Kanana Open License, which allows commercial use under its terms.

Kakao calls embeddings the foundation for its personalized AI agents. If you build Korean or bilingual search, RAG or recommendation systems, this release is worth a look.

What Kakao released

According to Kakao's press release and the Hugging Face model card:

Model Parameters Full embedding size Max input
kanana-2-3b-embedding 3B 2,560 4,096 tokens
kanana-2-1.3b-embedding 1.3B 1,280 4,096 tokens
  • Both models are built on Kakao's Kanana-2 small language models (SLMs), converted to bidirectional attention with mean pooling.
  • Kakao trained the 3B model first, then built the 1.3B model with knowledge distillation, where a larger model teaches a smaller one.
  • Both support Matryoshka representation learning (MRL). You can truncate embeddings to 1,024, 512, 256 or 128 dimensions to cut storage and search costs.
  • Kakao says the models are meant for semantic search, retrieval-augmented generation (RAG) and text similarity in Korean and English.

This follows the four Kanana small language models Kakao open-sourced in July.

How they score (Kakao's own benchmarks)

These are Kakao's numbers. The model card says every model was run with MTEB "under the same conditions" and "without any task prompt." We haven't reproduced them.

At full embedding size, Kakao reports:

  • Ko-Retrieval (average of six Korean retrieval tasks): kanana-2-3b-embedding 0.8211, kanana-2-1.3b-embedding 0.7958. By comparison, NVIDIA's llama-nemotron-embed-1b-v2 scores 0.7985 and Qwen3-Embedding-4B 0.7837.
  • nano-BEIR (Korean): 3B model 0.6088, against Qwen3-Embedding-4B at 0.5522.
  • MTEB (English, v2): 3B model 0.6317. On this benchmark, Qwen3-Embedding-4B scores higher at 0.6661.

Truncation costs some accuracy. At 128 dimensions, the 3B model's Ko-Retrieval score drops to 0.7488, still above Qwen3-Embedding-4B's 0.6854 at that size, per the model card.

Kakao's press release also says the models are "tens of times" more efficient in memory use and inference speed than its previous Kanana embedding models. Kakao didn't publish figures for that claim in the sources we checked, so treat it as unverified.

The everyday pitch: "that quiet restaurant"

Kakao's press release gives a simple example. A user looking for "the quiet restaurant we talked about before" may not remember its name. An embedding model can still surface past information with a similar meaning, such as "a good place to talk." That's what Kakao means by personal agents connecting the scattered context of daily life, a theme it says it highlighted at its if(kakao)26 conference.

License: read before you ship

The weights use the Kanana Open License (the model card links the LICENSE file). Kakao says it allows research, development and commercial use under the license's conditions. Read the full license before you put these models into a commercial product. We didn't review every clause.

How to try it

The model card lists sentence-transformers support and deployment notes, including Matryoshka truncation. Both models are on Hugging Face:

What's confirmed / not confirmed

Confirmed (Kakao press release and official model cards): - Two models (3B and 1.3B) were released on Hugging Face on Oct 29, 2026. - Embedding sizes are 2,560 and 1,280, with MRL truncation to 1,024, 512, 256 and 128, and a 4,096-token input limit. - The models were built from Kanana-2 SLMs; the 1.3B model was made by distillation. - They're released under the Kanana Open License, with commercial use allowed under its terms. - The benchmark figures above appear in Kakao's model card.

Not confirmed / company claims only: - The "tens of times" efficiency gain over older Kanana embedding models (no figures in the sources we checked). - Independent benchmark results. All scores are Kakao's own. - Which Kakao products will use these models, and when.

Sources

Related

Up: AI