Naver Also Faces Controversy Over 'Independent AI Project' Eligibility: "Adopted Chinese Qwen Vision Encoder"

Naver: "Efficient Optimization of the Entire System Considered"
"Independent Development Areas and Encoder Areas Should Be Distinguished"
Potential Issues If License Is Revoked

Following Upstage, Naver Cloud has also become embroiled in a controversy over its eligibility to participate in a government project aimed at building artificial intelligence (AI) models using domestic technology.

According to the information and communications technology (ICT) industry on January 6, there have been claims that Naver Cloud's HyperCLOVA X Seed 32B Sync model shows high cosine similarity and Pearson correlation coefficients between its vision encoder weights and those of Alibaba's Qwen 2.4 language model from China. A vision encoder is responsible for extracting visual information from images and videos and converting it into a form that a language model can understand.

Naver Cloud acknowledged the use of Chinese open source, explaining, "For this model, we strategically adopted a proven external encoder, taking into account compatibility with the global technology ecosystem and efficient optimization of the entire system."

The company added, "This was not due to a lack of technological independence, but rather a high-level engineering decision to enhance the completeness and stability of the overall model by utilizing already standardized high-performance modules."

Some members of the developer community have expressed surprise that Naver Cloud borrowed the vision encoder-a core function of multimodal AI that processes various types of information such as visual and auditory data-from Qwen, and that not just the dataset but even the weights were utilized.

In response, Naver Cloud stated, "The vision encoder serves as an optic nerve, converting visual information into signals that the model can understand," and emphasized, "Naver possesses its own unique visual technology capabilities, such as VUClip."

Yuwon Kim, CEO of Naver Cloud, Yonhap News Agency

A Naver Cloud representative said, "There should be a clear distinction between areas that must be independently developed, such as generative AI, and areas like encoders that do not require retraining," adding, "The development of AI technology is a process of advancing by adding our own unique value to the 'shoulders of giants,' namely, the globally proven big tech companies."

The company further explained, "While we respect the technological foundations already validated by academia, we have added advanced additional training and Naver's own optimization technologies to ensure that the model can accurately recognize the Korean cultural context."

However, if Qwen, which provided the open source, withdraws permission for its use, it remains uncertain whether the model can continue to be maintained.

The government's intent in developing an independent AI foundation model was to prepare for potential problems that could arise if U.S. or Chinese big tech companies suddenly increase the usage fees for AI foundation models or abruptly revoke usage licenses.

When launching this project, the Ministry of Science and ICT stipulated that the model should be a domestic one, designed and pre-trained from scratch, and not a derivative model developed by fine-tuning overseas models. The ministry also set as a condition that there should be no licensing issues with other companies' models.

When five elite teams were selected last August, the ministry stated, "All elite teams have proven their capabilities in developing high-level AI models and are evaluated as having pursued the essence of sovereign AI by starting from scratch."