As the use of downloadable open weight AI models in academic research grows globally, nearly half of these applications are centered around Alibaba's Qwen model in China. Analysts suggest that many researchers are opting for domestic models rather than prioritizing openness.
On August 20, the IT industry reported that researcher Zachary Okun-Dernivine from the University of Stuttgart's Institute for Social Science Research published a paper titled "Who Actually Uses Open Weight Models, and How is Their Usage Landscape Changing with China at the Center?" on arXiv.
The paper analyzed 157,446 studies that utilized AI models as experimental tools, selected from a total of 21,338,177 papers indexed in the S2ORC academic database.
The analysis revealed that in the first half of 2026, 44% of the 16,125 papers using a single model type employed open weight models, nearly equal to the 44.2% that used proprietary models from the GPT series. A significant concern is the composition within the open weight category. Among the papers that chose open weight models, nearly half, or 49.9%, utilized Qwen. When considering all Chinese open weight models, this figure rises to 60.5%, indicating a heavy reliance on a specific model from a particular country.
During the same period, Meta's LLaMA accounted for 8.6%, while MistralAI's Mistral was at just 0.8%. Both models saw a significant decline in usage compared to two years ago. The paper notes that despite LLaMA and Mistral releasing their weights, they have been overshadowed in academia, while Qwen and DeepSpeed have gained traction due to their competitive performance, lower costs, and rapid release cycles for new models. The mere act of releasing weights is more of an industrial strategy for market expansion and cannot solely explain their adoption.
The study also statistically examined the disparities in model selection based on the researcher's country. Researchers affiliated with Chinese institutions were 2.23 times more likely to choose open weight models than their counterparts from other nations.
However, the paper argues that this gap cannot be interpreted as a preference for openness among Chinese researchers. The likelihood of a researcher from a Chinese institution selecting an open weight model from a non-Chinese source (like LLaMA or Mistral) was only 15%, lower than the 18.5% for researchers not connected to China. This suggests that the concentration of open weight usage in Chinese academia is more about prioritizing domestically developed models rather than a preference for openness itself. From 2023 to the first half of 2026, the adoption rate of open weight models increased from 12.8% to 35.8%, with Chinese researchers contributing 60.2% of this growth.
The research also highlighted issues with data completeness. The proportion of papers with verifiable author affiliations was 77.1% overall, but this dropped to 65.9% for papers using open weight models and further to 49.6% for those using Qwen. This indicates a clear trend where newer papers, those utilizing open weight models, and those authored by Chinese researchers are more likely to lack affiliation information. The research team did not shy away from these limitations and included sensitivity analyses to correct for missing data.
In the competitive landscape of market expansion centered around domestic models, South Korea's position appears minimal. The paper's classification system grouped countries into U.S. and China categories, leaving South Korea lumped into a "non-China" category without separate identification. While there are models in South Korea, such as LG AI Research's ExaOne, Naver's HyperCLOVA X, Kakao's Kanana, and Upstage's Solar, that have released their weights, no research has yet measured their actual adoption in academia with this level of precision.
As the U.S. solidifies proprietary standards and China establishes open weight standards, South Korea finds itself unable to quantify how much its own models are utilized even within its academic circles. This raises concerns that discussions around independent AI foundation models need to evolve from merely asking whether a model has been created to whether that model is actually being used within the domestic research and industrial ecosystem.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.