Chinese artificial intelligence (AI) models are rapidly expanding their influence in South Korea's development sector. Notably, Alibaba's 'Qwen' has surpassed American big tech models in usage, indicating a broader impact of Chinese AI across the open AI ecosystem.
According to a report released on October 21 by Rockbridge Network Korea's Economic and Industrial Research Office, as of August, 49.3% of newly generated language models created on domestic Hugging Face accounts were based on Chinese models, compared to 35.9% for American models.
The growth of Chinese models has been steep. Their share in the domestic market was only 1.2% in 2023 but surged to 49.3% this year. In contrast, the share of American models dropped from 79.5% to 35.9% during the same period.
Leading this surge is Alibaba's Qwen, which accounted for 864 out of 1,845 new outputs this year, or 46.8%. Google's Gemma contributed 271 outputs (14.7%), Meta's LLaMA had 254 (13.8%), and LG AI Research's ExaOne produced 171 (9.3%).
This disparity suggests a significant gap between perceptions of Chinese AI in South Korea and its actual presence in development environments.
The report attributes Qwen's competitiveness to its openness and diverse model options, noting that it offers choices ranging from small to large models and has been trained in 119 languages.
However, Chinese AI has not yet dominated the domestic consumer market. DeepSeek, which topped the free app downloads in South Korea's app store last year, saw a sharp decline in users following privacy concerns and access restrictions by public institutions and companies. As of July, no Chinese AI chatbot services were among the top 10 monthly active users in South Korea. In contrast, users of Claude increased from 770,000 in February to 3.44 million in July, while Gemini's users grew from 440,000 to 9.47 million.
Companies remain cautious about openly using Chinese AI. Reasons cited for not adopting Chinese models include concerns over data leakage (80.2%), lack of transparency in personal data handling (67.7%), and worries about performance and quality (31.3%).
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.