South Korea’s AI contest has reached its serious stage. Upstage, SK Telecom and LG AI Research advanced after the second-round evaluation, while Motif Technologies was eliminated for ranking last among the four competing teams. The result leaves three teams competing for further development and a final selection process.
The Ministry of Science and ICT launched the Independent AI Foundation Model project in September with five teams. The government set a target for a Korean foundation model to reach about 95% of the performance delivered by frontier-class AI models — an ambitious benchmark, because apparently “good enough” was not dramatic enough.
Naver Cloud and NC AI left the contest after the first evaluation in January. An additional application process in February brought Motif Technologies into the program, creating a four-team field for the second round.
Technology Was Not the Only Test
The second-round results, announced Aug. 18, 2026, used three evaluation categories. Benchmark performance carried 40 points, expert evaluations carried 35 points and user evaluations carried 25 points.
The four teams averaged 22.5 points in benchmark testing, 28.8 points in expert reviews and 17.6 points in user evaluations. Those figures show a contest measuring more than raw model capability: technical results, expert judgment and practical user experience all shaped the outcome.
All four models were determined to have achieved technological independence. That gave every second-round participant a baseline achievement, but it did not erase the gap between technical strength and usefulness in real applications.
Ryu Je Myung, Second Vice Minister of Science and ICT, said different teams placed first in each evaluation category. The result suggests that no single model dominated every test, which makes the government’s selection process less about crowning one benchmark champion and more about deciding which systems can keep improving.
Motif Technologies shows why that distinction matters. “In the case of Motif Technologies, it demonstrated a world-class level from a technological standpoint, but it received lower evaluations than the other models in usability and practical application,” Ryu said.
That explanation also defines the contest’s pressure point. A model can reach a world-class level from a technological standpoint and still lose ground when users and experts judge whether it works well outside the test environment.
More GPUs and a New Support Structure
The three remaining teams will continue developing their AI models before a third evaluation scheduled for early 2027. The final two teams are expected to be selected after that evaluation, turning the current three-team field into the last stage of the program.
The government also plans to expand computing support by increasing Nvidia B200 GPUs from about 768 to roughly 1,000. The first half of this year included about 768 Nvidia B200 GPUs, and the planned increase gives the remaining teams more hardware for model development.
That expansion comes with a review of the program’s support system. The government plans to announce additional details later, while Ryu said the Ministry of Science and ICT and participating companies have discussed whether the current structure can keep pace with global frontier AI models.
The discussion points to a broader change in the project’s purpose. South Korea is not only evaluating which team performs best under the current rules; it is also reconsidering whether those rules and resources can support competition at the level of frontier companies’ models.
There is a shared understanding that the program should be reorganized to allow that level of competition. The next evaluation will therefore test more than the progress of three models — it will test whether the contest itself can produce the conditions needed for a Korean foundation model to approach its 95% target.
Based on
