South Korea's "self-reliant AI" cannot bypass Chinese models.

South Korean private think tank Rockbridge Network Korea released a survey report titled "How Chinese Artificial Intelligence Is Used in South Korea" in late October, authored by Choi Sung-jin, a professor of business administration at Hanyang University.

The report shows that back in 2023, Chinese AI models drew almost no attention in South Korea. However, between January and August 2026, nearly 50% of the new AI models added in South Korea were derived from Chinese AI.

Furthermore, this trend is intensifying. Choi Sung-jin believes that this figure will reach 70%-90% by 2027.

01

South Korean AI Can No Longer Function Without Chinese Models

Hugging Face does not mark nationalities on its platform, so users who want to find "South Korean developers" have to search for them manually. The research team wrote a set of programs that scanned all accounts on the platform using Korean tags and keywords, and finally locked 1993 South Korean accounts and 7939 models they had publicly released over the years.

The new models added between January and August 2026 are categorized by base framework: 864 Qwen models, accounting for 46.8%; 271 Google Gemma models, accounting for 14.7%; 254 Meta Llama models, accounting for 13.8%; 171 LG EXAONE models, accounting for 9.3%; 49 Kakao Kanana models, accounting for 2.7%; 19 DeepSeek models, accounting for 1.0%. The remaining 217 models account for 11.8%.

The most interesting part is that these 864 Qwen-based models come from 164 accounts, which means each account uploaded at least 5 models. A total of 1845 new models were added between January and August 2026. On average, nearly half of South Korean Hugging Face accounts have uploaded models based on Qwen.

But that is not the most interesting point. Choi Sung-jin sent questionnaires to 102 small and medium-sized enterprises, one of which asked "Do you use Chinese AI services in your work?"

None of the enterprises stated that they were using Chinese models.

This result did not convince Choi Sung-jin, so he asked these enterprises to present their models. As a result, 5 enterprises immediately corrected their statements and admitted that they were indeed using Chinese models.

In addition, 45.1% of the surveyed enterprises said that it was the first time they had heard of Qwen.

Choi Sung-jin said that a considerable number of respondents in the questionnaire had nothing to do with development work, so they had no idea what models their companies were running. Coupled with the stereotypes about Chinese AI, they were unwilling to tell the truth.

Later, Choi Sung-jin conducted a survey on Model Fingerprinting of the products of these companies, and found that far more than 5 enterprises used Chinese models as the base framework, and even the vast majority were using Chinese models.

This time last year, DeepSeek was the most widely used model in South Korea. But this year, only 19 models use DeepSeek as the base framework.

Choi Sung-jin believes that Qwen's main advantage lies in engineering. The smallest Qwen model is only a few billion parameters, for example, Qwen2 has a 0.5B version, and Qwen3 has a 0.6B version. The largest Qwen3.8-Max has 2.4T parameters, which basically achieves full-size coverage.

Choi Sung-jin also said that Qwen supports 119 languages, its open source license is mainly Apache 2.0, and it is equipped with a complete set of fine-tuning tools, quantized versions and community cases.

All these reasons combined make Qwen occupy the AI market in South Korea.

Price is another factor. Choi Sung-jin stated in the report that among models with the same performance, the unit output price of Chinese models is about 1/34 of that of US models. Even after running a full set of evaluations, the cost of Chinese models is only 1/4 of that of US models.

This is especially true for services with concentrated users and long single dialogues, such as AI character chat and high-concurrency customer service. In this type of business, model calling costs directly determine gross profit, so low price and stability are more important than performance.

What about South Korea's own models?

Among the 14.9% local market share, LG's EXAONE accounts for 9.3%, and Kakao's Kanana accounts for 2.7%.

But the problem does not lie in model performance, but in model availability.

EXAONE 4.0 (32B) uses the "EXAONE AI Model License 1.2 - NC".

NC stands for non-commercial use, which prohibits commercial use, and even the outputs generated using it cannot be sold. Developers using it to make products will be in breach of the agreement. It was not until K-EXAONE 2.0 (750B MoE) was released on July 31 this year that LG replaced the license with Apache 2.0.

However, domestic open source models such as Qwen and DeepSeek allow commercial use from the very beginning, and developers can fine-tune them and directly apply them to enterprise scenarios.

In January this year, South Korea launched the "Autonomous AI Foundation Model" project, and the first round of evaluation results were announced on January 15. Two seed players, NAVER Cloud and NC AI, were both eliminated.

The reason for NAVER Cloud's elimination is that its originality did not meet the standards. The submitted HyperCLOVA X SEED 32B Sync multimodal model directly reused the visual encoder weights of Qwen2.5, without even initial retraining. NC AI did not even reach the passing score.

In fact, the passing line for South Korea's "Autonomous AI Foundation Model" project is not high, which is roughly the level of Qwen3-235B. Combined with the timeline, the domestic models such as DeepSeek V3.2 and Kimi K2.5 were ahead of Qwen3-235B at that time.

Qwen3-235B is not even a flagship product within the Qwen family. After all, the Qwen3 family has Qwen3-Max, which is reported to have more than 1T parameters, and its performance is naturally much higher than that of Qwen3-235B.

02

Chinese AI Has Become the Global Base Framework

South Korea is not an isolated case. The fact that Chinese models have become global base frameworks is no longer surprising.

Hugging Face's own "State of Open Models" summer report shows that Qwen already has 151,448 derivative models on the platform, which is 2.6 times the total of Meta's models and 4.7 times the number of Llama series repositories. In the first seven months of 2026, an average of 180 to 210 new Qwen derivative models were added every day, and Qwen was downloaded about 2.045 billion times this year.

The report also points out that among the models of major Chinese AI manufacturers downloaded this year, the proportion of models with more than 70B parameters is as follows: MiniMax is almost 100%, Moonshot AI is 88%, DeepSeek is 55%, and Zhipu AI is 39%.

The situation in the United States is almost the opposite: Google, Microsoft, and IBM Granite are close to zero, Nvidia is 14%, and Meta is 9%.

Users choose large-sized models when downloading Chinese models, and choose the lightest ones when downloading US models.

In terms of download share, Chinese open-weight models account for 41% of the platform, and US models account for 36.5%.

In addition to developers on Hugging Face, many companies in the United States have also begun to use Chinese models. Cursor, an AI programming company, released Composer 2 in March this year, which uses Moonshot AI's Kimi K2.5 as the training base framework. Brian Chesky, CEO of Airbnb, said that their customer service system uses 13 models in total, and Qwen accounts for the largest proportion.

So there is nothing surprising about the current situation in South Korea, it is just a microcosm of the whole world.

Models such as GPT, Claude and Gemini are all closed-source, and can only be used by calling APIs. The open-weight track has been fully occupied by Chinese models.

Among the 178 Chinese open models with more than 20B parameters, 59% use the Apache 2.0 license, and 22% use the MIT license. For models of the same scale, only 29% of US models use Apache or MIT licenses, 41% use custom terms, and 30% do not specify any terms at all. In other words, Chinese models can be freely fine-tuned, while the vast majority of US open source models cannot be used for commercial purposes.

Openness itself is a pricing strategy.

The more complete the model size range, the more people are willing to fine-tune it; the more derivative models there are, the more tools, tutorials and quantized versions will be available; the next batch of people will still choose it as the base framework.

The GGUF version of Qwen that can run locally is downloaded 39.6 million times per month, which is twice that of Gemma and more than five times that of Llama. In fact, the number of GGUF repositories for the Llama series is slightly higher than that of Qwen.

Ecological advantage is never won by parameters, but by "allowing others to get things done with it".

In South Korea, in addition to continuing to promote the "Autonomous AI Foundation Model" project, a new 4.7 trillion won (about 3.5 billion US dollars) "Frontier AI" plan was added in October, which is scheduled to start in March 2027. 3.9 trillion won of the fund will be used to purchase about 10,000 Nvidia Vera Rubin GPUs, 800 billion won for training data, and another 25 billion won to recruit 20 overseas researchers.

The latest super-large models from several top Chinese laboratories have also begun to add non-commercial terms and revenue sharing requirements.

There is no free lunch in the world. As the AI industry gradually matures, charging appropriate fees is the basic condition for the healthy operation of enterprises.

This article is from the WeChat official account "Face AI" (ID: faceaibang), author: Miao Zheng, editor: Wang Jing, published with authorization from 36Kr.