[Alpha Biz= Paul Lee] Controversy has emerged regarding whether Naver Cloud’s HyperCLOVA X foundation AI model was built entirely from scratch.
According to sources in the information and communications technology (ICT) industry on January 5, the HyperCLOVA X Seed 32B Sync model appears to exhibit high cosine similarity and Pearson correlation coefficients in its vision encoder weights compared with China’s Alibaba Qwen 2.4 language model.
Cosine similarity and Pearson correlation coefficients are commonly used metrics to assess the similarity between datasets. Cosine similarity evaluates directional similarity, while Pearson correlation measures how closely the data distributions align.
The high similarity metrics between HyperCLOVA X and Qwen 2.4 could suggest partial usage of Qwen’s vision encoder and weights.
Naver Cloud acknowledged the use of Chinese open-source components in its model. A company spokesperson stated:
"In this model, we strategically adopted verified external encoders to ensure compatibility with the global technology ecosystem and optimize the overall system efficiently. This decision was made from an engineering perspective to enhance the model’s completeness and stability, rather than due to a lack of technical independence."
However, Naver Cloud strongly denied claims that the foundation model was not built from scratch.
Alphabiz Reporter Paul Lee(hoondork1977@alphabiz.co.kr)
