Chinese models underpinned 49.3% of Korea-linked derivative AI releases on Hugging Face through August 2026, a study reported by The Korea Herald found, highlighting a challenge for Seoul's domestic model effort.
South Korea's Ministry of Science and ICT said in August that Upstage, SK Telecom and LG AI Research advanced to the next stage of its Sovereign AI Foundation Model project.
Hanyang University business professor Choi Seong-jin examined 1,845 derivative model releases from Korea-linked accounts between January and August 2026 in an Oct. 16 study for Rockbridge Network Korea. U.S. foundation models supplied the base for 35.9% of those releases and Korean models for 14.9%.
Alibaba's Qwen was the base for 864 releases, or 46.8% of the total. Among Korean models, LG AI Research's Exaone accounted for 171 releases, or 9.3%, while Kakao's Kanana accounted for 49, or 2.7%.
The study put the Chinese-model share at 1.2% in 2023 and 22% in 2025. It linked Qwen's appeal to its developer tools, range of sizes and support for 119 languages.
The Hugging Face releases do not measure AI use across South Korea. No Chinese chatbot appeared among the country's 10 most-used AI apps in July. The researchers also found no references to Chinese models in about 1,900 public GitHub repositories from 36 major Korean companies, but noted that downloaded models can be used without accessing a Chinese provider's online service.
In a survey of 102 small and midsize businesses, 80.2% identified data leaks as a concern, while 31.3% cited model performance and quality. The report warned that developers relying heavily on one provider could face changes to its licensing terms or an end to new model releases. Choi wrote that Korean alternatives need licensing terms, size options and customization tools that work for developers.
