Conceptual diagram of AI Photo: VCG

Chinese-developed artificial intelligence large-language models are gaining ground among overseas developers, with a recent South Korean study finding that nearly half of the AI models released by local developers on Hugging Face this year were built on Chinese models, outpacing those based on US and South Korean technology.

The findings point to a broader shift in AI contest, as cost, flexibility and ease of customization gain importance, in addition to model performance, Chinese analysts said.

Of the 1,845 models published by South Korea-based accounts on Hugging Face between January and August, 49.3 percent were based on Chinese models, compared with 35.9 percent on US models and 14.9 percent on South Korean models, The Korea Times reported on Sunday, citing a study by think tank Rockbridge Network Korea.

Alibaba's Qwen led the field, accounting for 864 models, or 46.8 percent of the total, far ahead of Google's Gemma at 14.7 percent and Meta's Llama at 13.8 percent, as well as South Korea's LG AI Research EXAONE at 9.3 percent and Kakao's Kanana at 2.7 percent.

The figures underscore how quickly the balance has shifted. Chinese models accounted for only 1.2 percent of new releases in 2023 and 22 percent in 2025, before reaching 49.3 percent in the first eight months of this year. Over the same period, the share based on US models fell from 79.5 percent to 35.9 percent, The Chosun Daily reported.

The findings suggest that Chinese open-weight models had moved beyond being merely usable alternatives to becoming a default starting point for many developers, making it easier to move from prototypes to practical deployment, Chen Jing, a vice president of the Technology and Strategy Research Institute, told the Global Times on Monday.

"Developers do not simply choose models based on their country of origin. They make engineering trade-offs among cost, performance, licensing and the speed of updates," Chen said, noting that Chinese models span lightweight versions for on-device applications and larger systems capable of handling coding, tool use and long-context tasks.

Alibaba said in August that its Qwen model family had surpassed 3 billion cumulative downloads worldwide, with more than 300,000 derivative models created. More broadly, Chinese open-source AI models had recorded more than 10 billion cumulative downloads globally as of July, according to China's Ministry of Industry and Information Technology, underscoring their growing reach among developers worldwide.

The Korean think tank report cited low costs and a wide range of models as key attractions for South Korean developers. Unlike closed models whose underlying weights are not publicly available, open-weight models such as Qwen can be downloaded, modified and fine-tuned for specific tasks without necessarily relying on cloud-based AI services.

The cost gap can be substantial. Chinese models' output costs were about one-34th of those top-tier models, although the advantage narrowed to roughly fourfold in a full benchmark run, according to the report. 

Beyond cost advantages, Chen said, the strength of Chinese models increasingly lies in the development ecosystem built around them. "Documentation, software tools and community support make it easier to customize models and bring applications into practical use, lowering the barriers for smaller teams moving from experimentation to deployment."

The appeal is also translating into practical business operation. The Korea Herald cited an official at a South Korean AI startup as saying the company mainly used Chinese models such as Qwen and DeepSeek for services with high user traffic, where operating costs were a key consideration.

Yet the practical advantages of foreign models do not remove South Korea's concerns about tech autonomy. The country is pursuing homegrown alternatives through a government-backed sovereign AI program, which has received around 200 billion won ($151 million) in support, The Korea Herald reported.

Chen acknowledged such concerns but said the real risk lies in becoming locked into some closed systems that are difficult to replace, rather than in using Chinese models. Open-weight models can be downloaded, fine-tuned and migrated, giving developers the flexibility to choose alternatives rather than rely on a single provider, he said.

"The competition is shifting from model rankings toward ecosystems," Chen said, adding that the growing adoption of Chinese LLMs overseas could lower barriers for developers and give China a greater role in shaping global AI development practices and technical standards, contributing to a more diverse global AI landscape.