[Alpha Biz= Kim Jisun] Naver is moving to replace previously used Chinese technology in its artificial intelligence (AI) foundation model with in-house developed components, aiming to address concerns over technological independence and strengthen competitiveness tailored to Korean language and culture.

According to industry sources on February 17, Naver Cloud recently completed the development of its own vision encoder and has begun integrating it into its multimodal AI models. A vision encoder converts visual inputs such as images and videos into numerical representations that AI systems can process, serving as a core component in models that handle text, image, audio, and video data.

With the adoption of its proprietary encoder, Naver aims to build its AI systems entirely “from scratch,” covering the full development cycle from training to deployment.

The move follows controversy surrounding Naver’s earlier use of components from Alibaba’s open-weight AI model Qwen 2.5. The issue raised questions about the originality of Naver’s model and led to its exclusion from the first round of a government-led foundation model project in January.

Naver said its newly developed vision encoder significantly improves upon its previous in-house technology (VUClip) and achieves performance comparable to leading global models, including Qwen. The model has been trained in Korean from the outset, enabling direct linkage between visual data and the Korean language without translation, and improving accuracy in handling culturally specific contexts such as geography and proper nouns.

It remains undecided whether the new encoder will replace existing components in previously released models, including open-source versions such as HyperCLOVA X Seed 32B Sync.

 

 

 

[ⓒ 알파경제. 무단전재-재배포 금지]