Kakao Open-Sources Kanana-2-vision-encoder Vision Encoders in 426 Million and 98 Million Parameter Versions
Kakao Open-Sources AI Image Recognition Models
Kakao has open-sourced two versions of its self-developed image recognition model Kanana-2-vision-encoder: a medium version with 426 million parameters and a tiny version with 98 million parameters, both now available on Hugging Face. Kakao says that across 42 benchmarks spanning nine major categories, including general visual question answering, English-Korean OCR, document understanding and object grounding, the two models achieved the highest performance compared with Google SigLIP 2, Meta Perception Encoder, Apple AIMv2 and Alibaba Qwen3.5-ViT. The tiny version outperforms competitors with more than 400 million parameters on some benchmarks, with particularly notable gaps in Korean recognition, document understanding and object grounding. During training, the use of the self-developed Kanana-2 language model improved training efficiency by about 4x, and architectural improvements for high-resolution image processing reduced latency by more than 60%; the models are released under the Kanana Open License, permitting research, development and commercial use.
Why read this
AI score 57
Kakao has open-sourced two vision encoders and provided a 42-benchmark comparison against models such as SigLIP 2 and Perception Encoder, which can be used to gauge the competitiveness of small-parameter models.
Topics
Body
A performance comparison between Kakao's newly released AI models and overseas AI models. (Image: Kakao)
PC본문_6월 Mouser (in Article)//
AD Manager | AD178089411676 //
// AD Manager | AD178089411676
Kakao has released two versions of its internally developed artificial intelligence image recognition model, Kanana-2-vision-encoder, as open source.
The models include a medium version with 426 million parameters and a tiny version with 98 million parameters. Both are available on Hugging Face, a global AI model-sharing platform.
Vision encoders analyze visual information in images, including objects, text, positions and shapes, and pass the information to language models.
The technology is used in multimodal AI applications that recognize document content and structure or answer user questions based on photographs.
Kakao evaluated the two models across 42 benchmarks in nine categories, including general visual question answering (VQA), English and Korean optical character recognition (OCR), document understanding and object localization within images.
According to Kakao, both models achieved the highest performance across the nine evaluation categories compared with major competing models, including Google's SigLIP 2, Meta's Perception Encoder, Apple's AIMv2 and Alibaba's Qwen3.5-ViT.
The tiny model, with fewer than 100 million parameters, outperformed some competing models with more than 400 million parameters on certain benchmarks.
Kakao said the performance gap was particularly significant in Korean text recognition, document understanding and object localization.
The company said it used its internally developed Kanana-2 language model during training to improve the vision encoders' training efficiency by approximately four times compared with the previous approach.
Kakao also said improvements to the high-resolution image processing architecture reduced processing latency by more than 60% compared with major competing models.
Both models are released under the Kanana Open License, which permits research, development and commercial use.
Kakao plans to support companies and researchers in South Korea and overseas in developing Korean-language document recognition and image-based AI services using the models.
The company also plans to gradually release the language, vision and audio processing components of its internally developed multimodal AI model, Kanana-o.