By examining 7,939 generative language model outputs published by 1,993 South Korean accounts on Hugging Face, researchers tracked how developer choices evolved between 2023 and 2026. While cumulative totals on open platforms historically leaned toward Western architectures, recent monthly data reveals a sharp shift toward Chinese artificial intelligence models among South Korean developers.
The analysis combines app usage statistics, GitHub and Hugging Face activity records, corporate surveys, and in-depth interviews to map out how Chinese AI technology is utilized in South Korea.
Shifting Foundations in New Model Development
Although cumulative totals on Hugging Face showed that US-based foundations outnumbered Chinese ones, 60% of those US outputs were created before 2024. When breaking down outputs by creation period, the report shows that the share of Chinese-based new outputs rose from 1.2% in 2023 to 7.2% in 2024, 22.0% in 2025, and 49.3% between January and August 2026. Monthly comparisons indicate that Chinese-based outputs first surpassed US-based ones in February 2026.
Within the January to August 2026 window of 1,845 new outputs, Alibaba’s Qwen led the expansion by accounting for 46.8% with 864 items. Gemma followed at 14.7% (271 items), Llama at 13.8% (254 items), and domestic EXAONE at 9.3% (171 items), while DeepSeek-based outputs stood at just 1.0% (19 items). Qwen-based outputs reached approximately 3.2 times the volume of Gemma and 3.4 times that of Llama. Developers favored Qwen for its variety of model sizes, multilingual support, rich development tools, and practical use cases.
Public Metrics and Surveys Show Low Chinese AI Use
The report highlights a significant gap between public app metrics, open repositories, and actual developer choices. In July 2026, no Chinese services appeared in the top ten monthly active user rankings for domestic AI chatbot apps, which used a baseline of 300,000 monthly users. Corporate surveys of 102 small and medium-sized enterprises showed that when directly asked if they used Chinese AI services, no respondents answered that they currently used them, though some answered affirmatively when asked about specific services.
Security and privacy concerns remained the primary reasons for avoiding Chinese models, with 80.2% citing data leaks and 67.7% pointing to the opacity of personal information processing, while only 31.3% mentioned performance and quality. The research notes that service blocking of Chinese cloud providers does not prevent developers from downloading open weights to run in their own environments, leaving open the question of how dependency on specific foreign model families will influence future domestic development ecosystems.
