Kakao open-sources two Kanana-2 embedding models on Hugging Face

Kakao released two Kanana-2-Embedding models as open source on Hugging Face on October 29, 2026, according to techm.kr. The release covers Kanana-2-3b-Embedding, with 3 billion parameters, and Kanana-2-1.3b-Embedding, with 1.3 billion parameters.

Both models use the Kanana Open License, which permits commercial use as well as research and development. Kakao built the 3B model by training its Kanana-2-SLM small language model for information retrieval, and produced the 1.3B version through knowledge distillation from the larger model.

The embedding models adopt MRL (Matryoshka Representation Learning), which lets developers choose the embedding vector size. Kakao did not disclose benchmark results for the embedding models in the release.

The vision encoders, which Kakao published on Hugging Face earlier under the same license, come in a 426-million-parameter Medium version and a 98-million-parameter Tiny version. Kakao evaluated them across 42 benchmarks in 9 areas, including visual question answering, English and Korean OCR, document understanding and localization, and says they outperformed Google's SigLIP 2, Meta's Perception Encoder, Apple's AIMv2 and Alibaba's Qwen3.5-ViT. Kakao also says the encoders cut processing latency by more than 60% versus major global vision encoders.

In July 2026, Kakao released four Kanana small language models as open source, and at If Kakao 26 the company highlighted embedding technology as the starting point for personalized AI agents.