[AI Prism] The use of Chinese products has increased rapidly in the development of domestic AI models.

Alibaba Qwen-based output 46.8%... Utilization of domestic models is in single digits

The response to the corporate survey line was ‘0’… Check dependencies in individual developer work

(Seoul = Yonhap News) Reporter Hayoung Kwon = Among the artificial intelligence (AI) model outputs released by Korean developers, the proportion based on Chinese models reached 49.3% this year.

Although the avoidance of Chinese AI was evident in corporate surveys and application (app) usage statistics, a contrary pattern was confirmed in the actual model development records of individual developers.

◇ High proportion of Chinese products in domestic AI model development

According to the report 'How is Chinese artificial intelligence being used in Korea' published by private think tank Rockbridge Network Korea on the 20th, of the 1,845 AI model outputs released by 1,993 Korean accounts on the global AI model sharing platform Hugging Face as of August this year, 49.3% were based on Chinese models.

The proportion of output based on the Chinese model has risen sharply over the past three years, rising from 1.2% in 2023 to 7.2% in 2024 and 22.0% in 2025, and soaring to 49.3% this year.

On the other hand, the proportion of output based on the US model fell significantly from 79.5% to 35.9% over the same period.

The most utilized base model was Alibaba's Qwen. Of the total outputs, 864 (46.8%) were based on Qwen, followed by Google's 'Gemma' with 271 (14.7%) and Meta's 'Rama' with 254 (13.8%). There were only 19 (1.0%) products based on Chinese deep seek.

The proportion of domestic companies using models was relatively low. LG AI Research Institute's 'ExaOne'-based output ranked 4th with 171 (9.3%), and Kakao [035720]'s 'Canana'-based output was 49 (2.7%).

◇ Chinese AI use revealed in developer work outside the organization

These results show that it is difficult to determine the current status of Chinese AI use in Korea through surveys of companies or AI service usage statistics alone.

According to the report, in a survey of 102 small and medium-sized companies asking whether they were currently using Chinese AI services, not a single company answered that they were using them, and as of July of this year, Chinese services were not included in the top 10 monthly active users of domestic AI chatbot apps.

However, when we analyzed the developers' actual work records, a different picture emerged.

As a result of analyzing about 1,900 GitHub repositories of 36 major domestic companies, there was not a single case that mentioned the Chinese model. However, when narrowing it down to repositories of individual developers who were not affiliated with companies, 14 out of 57 (24.6%) individual repositories identified as Korean developers dealt with Chinese models.

The use of Chinese AI, which is difficult to capture only through the official usage status at the corporate level or the usage rate of popular apps, occupies a significant portion in the model development process of individual developers.

The report said, "Simply blocking access to Chinese AI services does not mean that Chinese AI is not used, and it is difficult to greatly influence the choices of developers who download public models and operate them directly. There is a need for companies and researchers to disclose model information transparently and diversify by reducing dependence on models from specific countries or affiliates."