Nearly half of new Korean AI services based on Chinese foundation models

Korean developers published 1,845 AI models on Hugging Face between January and August, with nearly half based on Chinese foundation models, raising concerns about technological dependence. A Rockbridge Network Korea report found that 49.3 percent used Chinese models, compared with 35.9 percent based on U.S. models and 14.9 percent on domestic models. Qwen accounted for 46.8 percent of the total, while the report cited low costs and adaptability as reasons for Chinese models’ appeal. Industry officials warned that dependence could create security, supply-chain and business-disruption risks.

The share of Korean developers’ new models based on Chinese foundations rose from 1.2 percent in 2023 to 22 percent in 2025 before reaching 49.3 percent this year.

Qwen accounted for 864 models, or 46.8 percent of the total; Google’s Gemma represented 14.7 percent, Meta’s Llama 13.8 percent, LG AI Research’s EXAONE 9.3 percent and Kakao’s Kanana 2.7 percent.

The report said Chinese models’ output costs were about one-34th those of comparable top-tier models, but the advantage narrowed to about fourfold across a full benchmark run.

In a survey of 102 small and medium-sized companies, 80.2 percent cited potential data leaks as a concern about Chinese AI models, and 67.7 percent cited a lack of transparency in personal data handling.

Published Oct 25, 2026 3:22 pm KST

Growing reliance raises security concerns

People stand by a logo of Qwen Intelligence, produced by Alibaba Cloud, at the 5th Global Digital Trade Fair in Hangzhou in eastern China's Zhejiang province on Sept. 22. AFP-Yonhap

Korea's artificial intelligence (AI) development ecosystem is increasingly relying on Chinese foundation models, with data showing that nearly half of the models newly released by local developers this year were built on them, raising concerns that the country's AI industry could become technologically subordinate to China.

According to a recent report by think tank Rockbridge Network Korea, Sunday, 49.3 percent of the 1,845 models published by Korea-based accounts on Hugging Face between January and August were based on Chinese models, compared with 35.9 percent based on U.S. models and 14.9 percent on domestic models.

The share of new models built on Chinese foundations has been rapidly growing in past years, surging from 1.2 percent in 2023 to 22 percent in 2025, before nearly doubling this year.

Qwen was the most widely used foundation model, accounting for 46.8 percent of the total, or 864 models. Google's Gemma ranked second with 14.7 percent, followed by Meta's Llama with 13.8 percent. LG AI Research's EXAONE ranked fourth at 9.3 percent, while Kakao's Kanana accounted for 2.7 percent.

Popular AI foundation models, such as OpenAI's GPT, Anthropic's Claude and Google's Gemini, are closed models whose underlying weights are not publicly available. In contrast, Alibaba's Qwen and Google's Gemma allow developers to download and modify the models for their own purposes.

Such models, known as open-weight models, are easier to adapt for specific applications thanks to their publicly available weights, including through fine-tuning with Korean-language data, without relying on cloud-based AI services.

The report noted that low costs and a wide range of available models have made Chinese models attractive to Korean developers.

Alibaba's Qwen offers models ranging from small versions to large-scale systems and supports 119 languages. Its extensive ecosystem of derivative models, tools and applications also gives developers greater flexibility in adapting the technology.

The report found that Chinese models' output costs were about one-34th those of comparable top-tier models, although the cost advantage narrowed to about fourfold when measured across a full benchmark run.

“The rapid adoption of Chinese open-weight models, including Qwen, reflects their growing use in lightweight experiments and day-to-day development work,” an industry official said.

Participants try SK Telecom-led consortium's artificial intelligence (AI) model A.X K1 at a presentation event for the national AI foundation model project at Coex in southern Seoul, Dec. 30, 2025. Courtesy of SK Telecom

The growing use of Chinese models nevertheless raises concerns about technological dependence. Alibaba offers both AI models and cloud infrastructure, potentially making it harder for developers to move away from its ecosystem as Qwen adoption grows.

“The biggest concern (in the rapid adoption) is not simply just information leaks, but the risk of technological dependence and supply-chain vulnerabilities,” the official said.

“As the United States and China compete for technological dominance, intensifying tensions could lead to sanctions on the use of models from a particular country or sudden changes in licensing policies, putting companies that have built services on those models at serious risk of business disruption.”

Security is another concern. In the report's survey on 102 small and medium-sized companies, 80.2 percent cited potential data leaks as a concern about Chinese AI models, while 67.7 percent pointed to a lack of transparency in personal data handling.

Industry officials cautioned that the opaque origins of training data in open-weight models could make it difficult for companies to meet strict requirements for AI safety, reliability and compliance.

“Rather than deepening reliance on external models in pursuit of short-term cost efficiency, Korea needs a two-track strategy that builds its own secure and reliable foundation models to safeguard technological sovereignty,” the official said. “Building good models is important, but building a strong ecosystem together is even more important.”

今年1月至8月，韩国开发者在Hugging Face平台上发布了1845个AI模型，其中近一半基于中国基础模型，引发了人们对技术依赖的担忧。Rockbridge Network Korea的一份报告显示，49.3%的模型使用了中国模型，相比之下，基于美国模型的占比为35.9%，基于韩国本土模型的占比仅为14.9%。Qwen的模型占总数的46.8%，报告指出，低成本和易用性是中国模型吸引用户的原因。业内人士警告称，这种依赖可能带来安全、供应链和业务中断等风险。

基于中国基础架构的韩国开发商新模型的份额从 2023 年的 1.2% 上升到 2025 年的 22%，今年更是达到了 49.3%。

Qwen 占总数的 864 个模型，即 46.8%；谷歌的 Gemma 占 14.7%，Meta 的 Llama 占 13.8%，LG AI Research 的 EXAONE 占 9.3%，Kakao 的 Kanana 占 2.7%。

报告称，中国车型的产出成本约为同类顶级车型的 1/34，但在完整的基准测试中，这一优势缩小到大约四倍。

在对 102 家中小企业进行的调查中，80.2% 的企业表示担心中国人工智能模型存在潜在的数据泄露风险，67.7% 的企业表示担心个人数据处理缺乏透明度。

发布于2026年10月25日下午3:22（韩国标准时间）

日益增长的依赖性引发安全担忧

9月22日，在浙江省杭州市举行的第五届全球数字交易会上，人们站在阿里云旗下产品“启文智能”的标识前。法新社-韩联社

韩国人工智能 (AI) 开发生态系统越来越依赖中国的基础模型，数据显示，今年韩国本土开发者新发布的模型中，近一半都是基于这些模型构建的，这引发了人们的担忧，即韩国的人工智能产业可能会在技术上从属于中国。

根据韩国智库 Rockbridge Network Sunday 最近的一份报告，1 月至 8 月期间，韩国账户在 Hugging Face 上发布的 1845 个模特中，49.3% 是以中国模特为原型，35.9% 是以美国模特为原型，14.9% 是以韩国本土模特为原型。

近年来，基于中国基础设施建设的新车型的份额迅速增长，从 2023 年的 1.2% 飙升至 2025 年的 22%，今年又几乎翻了一番。

Qwen 是使用最广泛的基础模型，占总数的 46.8%，即 864 个模型。谷歌的 Gemma 位居第二，占比 14.7%，Meta 的 Llama 紧随其后，占比 13.8%。LG AI Research 的 EXAONE 位列第四，占比 9.3%，而 Kakao 的 Kanana 占比 2.7%。

一些流行的AI基础模型，例如OpenAI的GPT、Anthropic的Claude和谷歌的Gemini，都是封闭模型，其底层权重并不公开。相比之下，阿里巴巴的Qwen和谷歌的Gemma允许开发者下载并修改模型以满足自身需求。

这种模型被称为开放权重模型，由于其权重是公开可用的，因此更容易针对特定应用进行调整，包括通过使用韩语数据进行微调，而无需依赖基于云的 AI 服务。

报告指出，低成本和种类繁多的车型使得中国车型对韩国开发商具有吸引力。

阿里巴巴旗下的Qwen平台提供从小型到大型系统的各种模型，并支持119种语言。其丰富的衍生模型、工具和应用生态系统也为开发者提供了更大的技术应用灵活性。

报告发现，中国车型的产出成本约为同类顶级车型的 1/34，不过，如果以完整的基准测试结果来衡量，成本优势缩小到大约四倍。

一位业内人士表示：“包括 Qwen 在内的中国轻量级机型迅速普及，反映出它们在轻量级实验和日常研发工作中日益广泛的应用。”

2025年12月30日，在首尔南部COEX会展中心举行的国家人工智能基金会模型项目展示活动上，参与者试用了由SK Telecom牵头的联盟开发的AI模型AX K1。图片由SK Telecom提供。

然而，中国模型的日益普及也引发了人们对技术依赖性的担忧。阿里巴巴既提供人工智能模型，也提供云基础设施，随着Qwen的普及，这可能会使开发者更难脱离其生态系统。

这位官员表示：“（快速采用新技术）最大的担忧不仅仅是信息泄露，而是技术依赖和供应链脆弱性的风险。”

“随着美国和中国争夺技术主导地位，日益紧张的局势可能导致对使用特定国家模型实施制裁，或许可政策发生突然变化，这将使基于这些模型构建服务的公司面临严重的业务中断风险。”

安全是另一个令人担忧的问题。该报告对102家中小型企业进行了调查，结果显示，80.2%的企业担心中国人工智能模型存在数据泄露风险，而67.7%的企业则指出个人数据处理缺乏透明度。

业内人士警告说，开放权重模型中训练数据的来源不透明，可能会使公司难以满足人工智能安全性、可靠性和合规性的严格要求。

这位官员表示：“韩国不应为了追求短期成本效益而加深对外部模式的依赖，而应采取双轨战略，构建自身安全可靠的基础模式，以维护技术主权。构建良好的模式固然重要，但共同构建强大的生态系统更为重要。”