Korean AI developers increasingly rely on Chinese open-weight foundation models

A report by Rockbridge Network Korea found that Korean developers published 1,845 AI models on Hugging Face from January to August, with 49.3% built on Chinese foundation models. This marks a sharp rise from previous years, while U.S.-based models comprised 35.9% and domestic models 14.9%. Alibaba's Qwen was the most popular, representing 46.8% of the total, with Google's Gemma and Meta's Llama trailing.

The appeal lies in the open-weight nature of these models, allowing developers to download, modify, and fine-tune them cost-effectively for Korean language tasks. However, officials caution that reliance on Chinese models may lead to data-leak risks, opaque training data, and potential sanctions or licensing changes that could disrupt businesses. They call for a dual strategy to develop secure, homegrown foundation models to preserve technological sovereignty.