Kakao Releases Two Open Source Kanana Vision Encoders

Kakao has released two versions of its Kanana-2-vision-encoder as open source, reports The Elec. The medium model has 426 million parameters, while the tiny model has 98 million, and both are available on Hugging Face.

Kakao tested the models across 42 benchmarks covering visual question answering, English and Korean text recognition, document understanding and object localization. The company reported that both models led nine evaluation categories against models including Google's SigLIP 2, Meta's Perception Encoder, Apple's AIMv2 and Alibaba's Qwen3.5-ViT.

Kakao says using its Kanana-2 language model in training made it about four times more efficient than its previous approach, and that changes to high resolution image processing cut latency by more than 60 percent compared with major competing models. The Kanana Open License permits research, development and commercial use.