Kakao releases Kanana-2 vision encoder AI models as open source

Kakao has released two versions of its proprietary AI image recognition technology, known as the ‘Kanana-2-vision-encoder’, as open source via the Hugging Face platform. The release includes a ‘medium’ model with 426 million parameters and a ‘tiny’ model with 98 million parameters.

According to Kakao’s internal evaluations across 42 benchmark items, the models outperformed several major global competitors, including Google’s SigLIP 2, Meta’s Perception Encoder, Apple’s AIMv2, and Alibaba’s Qwen3.5-ViT. The company noted significant performance advantages in Korean optical character recognition (OCR), document understanding, and spatial awareness within images.

To improve efficiency, Kakao utilized its own ‘Kanana-2’ language model during the training process, reportedly increasing training efficiency by approximately four times. Additionally, the company implemented high-resolution image processing improvements that reduced latency by more than 60% compared to major global models. The models are released under the ‘Kanana Open License’, which permits both research and commercial use. Kakao plans to continue releasing components for its upcoming multimodal ‘Kanana-o’ model, which aims to process text, images, and voice.