Kakao releases the visual components that read images for language models
Kakao has released two Kanana-2 vision encoders on Hugging Face under a licence allowing research, development and commercial use. The components extract objects, text and layout from images for a language model to interpret. The release follows an earlier conference preview and includes a compact model developed with on-device use in mind.
Artificial Intelligence··Evening
Image features pass into the language model
Kakao has released the medium and tiny versions of Kanana-2-vision-encoder on Hugging Face, a platform for sharing AI models. The technology company announced the release on 21 October after previewing it at its 13 October conference. A vision encoder supplies the visual part of a system that processes different kinds of information: it extracts objects, lettering, positions and shapes from an image and passes those features to a language model. The medium encoder contains 426 million parameters and tiny contains 98 million.[1], [2]
Small lettering and document layouts guide the training
The language model interprets those visual features alongside the user's question to produce an answer. Examples include reading the food and prices on a menu, understanding a document's contents and page layout, and describing a situation in a photograph. Kakao says its training methods are designed to preserve small lettering and fine visual details in complex documents. It uses its own Kanana-2 language model in encoder training and applies knowledge distillation and model compression. The tiny version was developed with use directly on a device in mind.[2]
Kakao evaluated the models on 42 benchmarks across nine areas. The tests include visual question answering, English and Korean text recognition, document understanding and identifying where things appear in an image. The reported comparisons with other encoders are the company's own evaluations. Comparisons include Google's SigLIP 2, Meta's Perception Encoder, Apple's AIMv2 and Alibaba's Qwen3.5-ViT.[1], [2]
The licence also permits commercial use
Both models use the Kanana Open License, which permits research, development and commercial use. They are components of Kakao's wider Kanana-o work, alongside language, visual and audio processing modules developed inside the company. Kakao says it plans to release additional components in stages. The current announcement concerns the two vision encoders; the wider system's remaining components are described as planned releases. Noh Byung-seok, who leads the company's unified foundation-model work, said Kakao intended to continue developing its models and sharing related technology.[1], [2]
