The company replaces Alibaba Qwen-based vision encoder with in-house version, aiming to end originality backlash
Headquarters building of Naver. (PHoto courtesy of Naver)
Naver will completely remove the controversial Chinese vision encoder used in its ‘Independent AI Foundation Model’ development project, opting instead to fully implement its own in-house vision encoder across its AI models. This move aims to resolve the long-standing controversy over the model’s originality and enhance its multimodal competitiveness, specifically tailored to the Korean cultural context.
According to the IT industry on April 17, Naver Cloud completed the development of its proprietary vision encoder early last month and has initiated the process of internalizing it for application across all future multimodal models. A vision encoder is a module that converts image and video information into a format that an AI can understand. In multimodal models, which comprehensively process text, image, audio, and video information, it functions as a kind of ‘optic nerve’.
Previously, Naver Cloud sparked controversy by partially borrowing the vision encoder and weights from Alibaba’s Qwen 2.5 model for its own multimodal model, ‘HyperCLOVA X SEED 32B Sync’.
The project was launched to reduce dependency on foreign big tech AI models and cultivate a uniquely Korean AI model. Naver’s use of Chinese technology drew heavy criticism as the project had championed the principle of building with its own technology from the earliest learning stages, a ‘From Scratch’ approach. At the time, Naver Cloud argued, “The vision encoder can be replaced at any time, and it is not a core area that is impossible to replace.”
With this development, Naver plans to put the originality controversy to rest and lay the foundation for technological independence. The new encoder is assessed to have performance on par with ‘Qwen’, a top-tier global model. Its standout strength is its visual intelligence, which accurately interprets the context of the Korean language and culture. The newly developed vision encoder is structured to directly link images with the Korean language, bypassing a separate translation process. For instance, while existing global encoders might recognize an image of Korea’s ‘Harubang’ simply as a ‘Statue’, Naver’s model immediately associates it with the Korean word ‘Harubang’.
A Naver Cloud official emphasized, “The key differentiator is that it is designed to directly connect the Korean language and images, enabling it to read the unique context of our culture without information distortion. When handling visual data that includes Korean geography, culture, or proper nouns, it will show an unrivaled accuracy that is on a different level from foreign models.”
Kim Eun-jin
댓글
0