Korea's national AI model project adrift as government questions own blueprint

Korea’s national artificial intelligence (AI) foundation model project is showing signs of running aground, as the government, which launched the program and pushed it toward a final evaluation despite controversies, has begun questioning its own blueprint.

Deputy Prime Minister and Minister of Science and ICT Bae Kyung-hoon wrote on Facebook on Tuesday night that the project will continue despite “various projections and speculation,” adding, “I believe Korea must also pursue frontier AI at the same time.”

“The gap between countries that have frontier AI and those that do not could become far wider than it is today,” Bae wrote. “But pursuing frontier AI should not mean stopping the national foundation model project. Not all AI needed in industries has to be a massive frontier model. What is needed is AI optimized for each field after considering cost, speed, security and specialization.”

Bae’s post came after reports that the ICT ministry is conducting a comprehensive review of the project’s direction, including whether to proceed with its final-round evaluation. The ministry is also reportedly considering establishing a special purpose company (SPC) dedicated to developing a frontier AI model.

Launched in June 2025, the national AI foundation model project was designed to foster Korea’s sovereign AI capabilities. The project drew significant attention as a competition to develop Korea’s most capable homegrown AI model, with a target of achieving large language model performance equivalent to at least 95 percent of leading global frontier models.

The government initially selected five teams and has since narrowed the field to three — LG AI Research, SK Telecom and Upstage — following the second-round evaluation. The process has faced controversy as other prominent domestic AI developers, including Naver, Kakao and Motif Technologies, were eliminated after failing to meet evaluation criteria.

The government had originally planned to conduct the third evaluation by the end of this year and select two final teams. However, questions have lingered in the market over whether their models are competitive against the latest models from OpenAI, Anthropic and Google.

Bae himself said last month that the project’s evaluation system was “lagging behind global standards.” The government is also reportedly questioning whether the competition framework designed last year remains suitable for an AI landscape that is changing rapidly.

“The government fully recognizes the need to concentrate financial support on developing top-tier AI models,” the ministry said in a statement. “However, no decision has been made on whether to terminate the national AI foundation model project or establish an SPC for the development of top-tier AI models.”

Bae, in his Facebook post, appears to assure the project's continuity, saying its “third-round evaluation will be completed as planned.”

However, the government’s sudden push for a separate frontier AI initiative has created fresh uncertainty for the industry. After designing the foundation model project as a competition to develop Korea’s most capable AI model, the government is now considering a separate initiative to develop even more capable frontier models.

“While the national foundation model project has had positive effects, such as expanding the talent pool, there were concerns from the beginning that companies eliminated from the competition could be perceived as lacking technological capabilities,” an industry official said.

“If the goal is truly to develop a frontier model, the government needs to carefully consider how the competition should be structured. Given that companies have already committed significant resources to the foundation model project and produced results, any frontier AI project — if pursued — should be linked to the existing project by any means.”

Independent lawmaker Han Dong-hoon criticized the government’s review of the project, questioning its explanation that the AI landscape has changed and that the continued advancement of U.S. and Chinese models renders the foundation model project alone insufficient.

“U.S. and Chinese AI were already continuing to advance back then,” Han wrote in Facebook post on Wednesday. “What was lacking was our strategy.”