Kakao open-sources Kanana-2 vision encoder models on Hugging Face
Kakao announced on October 22 that it has released two self-developed image recognition models, Kanana-2-vision-encoder, as open source on Hugging Face. The release includes a 426-million-parameter medium version and a 98-million-parameter tiny version, both under the Kanana Open License, which permits research and commercial use. Kakao says its internal evaluations showed the models outperforming several mainstream international vision encoders across visual question answering, English and Korean OCR, document understanding, grounding and fine-grained recognition, though no independent benchmark results were cited.
Kakao is opening core vision-encoder technology under a commercial-use license, which the company says is meant to lower the barrier for researchers and businesses adopting multimodal AI.

Kakao 於 10 月 22 日宣布,已將旗下自研的兩款影像辨識模型 Kanana-2-vision-encoder 釋出為開源模型,並上架至全球 AI 開發平台 Hugging Face。這次公開的版本包括 4.26 億參數的 medium 與 9,800 萬參數的 tiny,兩者皆採用可供研究與商業用途的 Kanana Open License。
這類視覺編碼器可視為多模態 AI 的「眼睛」,負責從影像中擷取物件、文字、位置與形狀等資訊,再交由語言模型進一步理解與回應。Kakao 表示,這類技術可應用於菜單與文件辨識、圖片內容理解,以及視覺問答等場景。
效能超越主流模型,訓練效率大幅躍升
根據 Kakao 的內部評測,兩款模型在涵蓋一般視覺問答、英文與韓文 OCR、文件理解、影像定位與細粒度辨識等多項基準測試中,整體表現優於多款國際主流視覺編碼器。
#Kakao#Kanana-2-vision-encoder#Kanana-2#Hugging Face#Kanana Open License