Seven of the 11 AI models deployed on South Korea’s Pan-Governmental Super-Sized AI Common Infrastructure were foreign-made as of September, according to data supplied by two ministries to a lawmaker, Chosun Ilbo reported on 5 October. The Ministry of Science and ICT said it had set no criteria during the platform’s development for assessing the country or institution behind a model.
Key points
- Of 31 models introduced to the infrastructure, 19 were foreign and 12 domestic as of September.
- Samsung SDS’s Fabrix and Naver Cloud’s CLOVA Studio for Gov are the two operational platforms.
- The model list includes Bge-M3, developed by an institution added to the US Entity List in March 2025.
- The ministry says providers select models and the government reviews their functionality, performance and security.
Seven models deployed, 31 introduced
The government has promoted the infrastructure since 2025 to let central and local agencies share AI models and GPU resources. Its stated aims include supporting administrative work and establishing sovereign AI for public-sector use. Separately, the government has pursued a homegrown foundation-model project intended to reduce dependence on overseas AI.
The figures supplied by the Ministry of the Interior and Safety and the Ministry of Science and ICT to Kim Jang-gyeom, a People Power Party lawmaker, distinguish models introduced to the infrastructure from those already deployed. The models not yet deployed can be activated at any time. The distinction matters for agencies choosing what to use: seven foreign models are already available among the 11 deployed, while the wider selection contains 19 foreign models among 31.
Kim contrasted that selection with the government’s treatment of Naver Cloud in a January evaluation of its homegrown foundation-model project. The government disqualified Naver Cloud for using Alibaba model weights, which it judged to lack originality. Kim argued that the common infrastructure used by public servants gave domestic models no corresponding preference.
Fabrix and CLOVA Studio for Gov
The infrastructure runs through Samsung SDS’s Fabrix and Naver Cloud’s CLOVA Studio for Gov. The largest model on Fabrix is Upstage’s Solar Open2 250B. Naver Cloud uses Nvidia’s Nemotron 550B. In those model names, B denotes billions of parameters.
Kim’s office raised a cost concern about Nemotron, saying its lengthy reasoning processes could require substantial computing resources and increase operating costs. GPU operating costs are factored into providers’ service fees paid by the government. The Ministry of Science and ICT has not disclosed token-based prices, flat-rate fees or the criteria used to calculate costs.
Bge-M3 and the ministry’s selection criteria
Among the 31 introduced models is Bge-M3, developed by the Beijing Academy of Artificial Intelligence. The US Department of Commerce added the academy to its Entity List in March 2025, Chosun Ilbo reported. The US cited its acquisition without authorisation of controlled items, including high-performance AI chips, and its work on large AI models and advanced chips for defence purposes.
The Ministry of Science and ICT described Bge-M3 as an embedding model, which turns text into numerical representations for computer processing, rather than a generative model. It said the model operates on a physically isolated administrative network intended to prevent data leaks. Embedding models can be used to retrieve documents, including laws and official papers, for systems that then generate an answer. Manipulation of such a model could affect which documents are retrieved, while checking its weights for deliberate tampering is practically impossible.
The ministry said providers chose models for their general utility, search performance, multilingual capability and ease of use. The government does not bar models on the basis of country, and reviews functionality, performance and security before introducing them, the newspaper reported. Kim questioned what that security review covered given the inclusion of a model developed by an institution on the US Entity List.
