Kakao open-sources two Kanana-2 vision encoder models
Kakao released Medium and Tiny Kanana-2 vision encoders under the Kanana Open License, which permits commercial use as well as research.
Kakao released two Kanana-2 vision encoder models, Medium and Tiny, as open source on Hugging Face on October 21, 2026, according to digitaltoday.co.kr.
The Medium model has 426 million parameters, while the Tiny model has 98 million. Kakao evaluated both on 42 benchmarks across 9 areas, including visual question answering, English and Korean OCR, document understanding and localization.
The company said its internal evaluation found both models outperformed comparison models including Google SigLIP 2, Meta Perception Encoder, Apple AIMv2 and Alibaba Qwen3.5-ViT. Kakao also said it used its own Kanana-2 language model to train the vision encoders, improving training efficiency about fourfold, and that high-resolution image handling cut processing latency by more than 60% compared with major global comparison models.
The models are published under the Kanana Open License, which allows commercial use as well as research and development. At its if(kakao)26 event on October 13, 2026, Kakao had announced plans to release the models.
Kakao said it plans to release other components of Kanana-o, its omni language model, sequentially.
Quick answers
What did Kakao release on Hugging Face?
Two Kanana-2 vision encoder models, Medium with 426 million parameters and Tiny with 98 million parameters, released on October 21, 2026.
Can the Kanana-2 vision encoders be used commercially?
Yes. They are released under the Kanana Open License, which allows commercial use as well as research and development.
What else is Kakao releasing?
Kakao said it plans to release other components of its omni language model Kanana-o sequentially.