Chinese Models Underpin Nearly Half of New Korean AI Releases

Chinese foundation models underpinned 49.3 percent of the 1,845 AI model derivatives published by Korea-linked Hugging Face accounts from January through August 2026, according to a Rockbridge Network Korea study reported by The Korea Herald. US models accounted for 35.9 percent, while Korean models accounted for 14.9 percent.

Alibaba Cloud's Qwen was the most widely used foundation model, supporting 864 derivatives, or 46.8 percent of the total. Google's Gemma accounted for 14.7 percent and Meta's Llama for 13.8 percent. LG AI Research's EXAONE represented 9.3 percent, while Kakao's Kanana represented 2.7 percent.

The share based on Chinese models rose from 1.2 percent in 2023 and 22 percent in 2025. The report cited Qwen's range of model sizes, support for 119 languages, available development tools and lower operating costs as factors in its use.

South Korea's government-backed sovereign AI program has received about 200 billion won, or $151 million, to develop domestic foundation models. Naver Cloud was removed from the program earlier this year after using Qwen components, which did not meet its technology independence requirements.