Alibaba is winning Korean developers through free open weights while Seoul's sovereign AI program demands models trained from scratch, a rule that removed Naver Cloud after it used Qwen weights.
Alibaba’s Qwen has become the most common starting point for AI models built in South Korea. Of 1,845 models published on Hugging Face, the main platform for sharing open AI models, by Korea-based accounts between January and August 2026, 864, or 46.8%, were derived from Qwen, according to a report by the private think tank Rockbridge Network Korea. The Korea Herald reported that the study was written by Choi Seong-jin, a business professor at Hanyang University.
Counting all Chinese model families, 49.3% of this year’s Korean models were built on Chinese base models, against 35.9% on US models and 14.9% on Korean ones. Chinese base models accounted for 1.2% of Korean derivatives in 2023 and 22% in 2025, before rising to 49.3% in the first eight months of 2026. Over the same span the US share fell from 79.5%, the Korea Herald reported.
The figures measure what developers build on, which is a different thing from what consumers use. No Chinese chatbot was among Korea’s 10 most-used AI apps in July, and the researchers found no references to Chinese models in about 1,900 public GitHub code repositories from 36 large Korean companies. "Simply blocking access to Chinese AI services does not mean Chinese AI is no longer being used," Choi told the Korea Herald.
Qwen leads US and Korean model families by a wide margin
Behind Qwen, Google’s Gemma accounted for 14.7% of derivatives and Meta’s Llama for 13.8%. The leading Korean family, LG AI Research’s EXAONE, ranked fourth with 171 derivatives, or 9.3%, and Kakao’s Kanana was sixth with 49, or 2.7%.
Price is part of the draw. The report found that Chinese models’ output costs were about one thirty-fourth of those of comparable top-tier models, although the gap narrowed to about four times when measured across a full benchmark run, the Korea Times reported. Because Qwen and Gemma publish their weights, developers can download them and retrain them on Korean-language data without relying on a foreign cloud service, Guancha noted in a Chinese-language report republished by The Paper.
Developers describe the choice in practical terms. "We mainly use Chinese models such as Qwen and DeepSeek for services with high user traffic," an official at a Korean AI startup told the Korea Herald. Choi said Korean open models need more than performance to compete: "They also need commercially viable licenses, a range of model sizes and the tools developers need to customize them."
Korean companies and regulators remain wary of Chinese models
The same report surveyed 102 small and midsized Korean businesses. Of those, 80.2% cited fears of data leaks from Chinese AI models and 67.7% pointed to a lack of transparency in how personal data is handled, according to the Korea Times.
An industry official quoted by the Korea Times said the larger risk was dependence. Because Alibaba sells both models and the cloud infrastructure to run them, developers may find it harder to leave its ecosystem, and US-China tensions "could lead to sanctions on the use of models from a particular country or sudden changes in licensing policies," the official said.
Seoul has already drawn a line in its own program. The Ministry of Science and ICT dropped Naver Cloud from the national AI foundation model project in January, saying it had not trained its model from newly initialized weights, according to a Korea Times report from January 15. Naver Cloud had used pre-trained vision encoders and weights from Qwen, according to a Korea Herald report from January 20. The government-backed program has about 200 billion won (US$151 million) in funding, and LG AI Research, SK Telecom and Upstage advanced to its next stage in August.
Korea’s response will test how far open weights travel
Guancha’s report closed with an earlier, undated statement from China’s Foreign Ministry that countries should promote open and inclusive AI development for the benefit of all humanity. Neither Alibaba nor the Korean government is quoted in the coverage responding to the report.
Qwen’s spread in Korea fits a wider pattern of Alibaba’s open models turning up in foreign systems, including a US federal website that briefly used Qwen for regulatory search in September. The industry official quoted by the Korea Times called for a two-track approach in which Korea builds its own secure and reliable foundation models to protect its technological sovereignty.
The measure to watch is the next round of Hugging Face counts. If EXAONE or the models from the sovereign AI program gain derivatives while Qwen’s share holds near 47%, Korean developers will be running two ecosystems side by side; if Qwen’s share keeps rising, the domestic program will have done little to change what Korean developers build on.