Naver will completely remove the Chinese vision encoder that had sparked controversy when it was included in its AI model during participation in the government-led “sovereign AI foundation model” project. The company plans to fully adopt a self-developed vision encoder across its AI models going forward.
Through the development of its own encoder, Naver is expected to address earlier controversy over its “from-scratch” approach and strengthen its “sovereign AI” strategy by fostering AI models better suited to the Korean language and culture in Asia.
According to the information technology (IT) industry on the 17th, Naver Cloud completed development of its in-house vision encoder early last month. It is currently working on internal integration to apply the encoder across all multimodal models it plans to develop in the future.
A vision encoder is a module that converts images and videos into formats that AI can understand. It functions as the visual intelligence component of multimodal models that process text, images, audio and video in an integrated manner.
The newly developed vision encoder by Naver Cloud shows significantly improved performance compared with its previous in-house encoder technology, VUClip. It is reportedly on par with global model encoders, which has proven utility in the open-source ecosystem.
Naver Cloud had previously faced controversy when it fine-tuned the vision encoder of Alibaba’s Qwen 2.5 model and largely adopted its audio encoder and weights in its omnimodal model, “HyperCLOVA X SEED 32B Sync,” during participation in the government-led project earlier this year.
The move drew criticism for contradicting the project’s principle of building models entirely with proprietary technology from the initial training stage.
However, it has not yet been decided whether the encoder in the already open-sourced “HyperCLOVA X SEED 32B Sync” model will be replaced with the newly introduced in-house encoder.
Naver Cloud’s new vision encoder is trained in Korean from the initial learning stage, enabling direct linkage between images and the Korean language without a separate translation process. It also minimizes information distortion by better understanding Korean culture and context.
A Naver Cloud official said, “When handling visual data that includes Korean geography, culture and proper nouns, the model is expected to demonstrate a level of accuracy distinct from foreign models.”