The startup Motif Technologies, which achieved the best performance among domestic firms in a global artificial intelligence (AI) evaluation, has been eliminated from the government’s competition to select “national flagship AI” models. Although Motif’s model technology itself was assessed to be at a world-class level, three other teams outperformed it in the overall evaluation, which also considered practical usability, potential for industrial field application, and contribution to the ecosystem. As a result, LG AI Research Institute, SK Telecom, and Upstage have advanced to the next stage and will compete early next year for the final two slots.

● AAII No. 1 Motif eliminated… “Technology at world-class level”

The Ministry of Science and ICT (MSIT) announced on the 18th that three teams—LG AI Research Institute, SK Telecom, and Upstage—had passed the second-stage evaluation of the Independent AI Foundation Model (Dokpamo) project. Dokpamo is a government program that provides support such as graphic processing units (GPUs) and data to build AI models led by domestic companies from architecture design through to training. In the first evaluation held in January, in which five teams competed, NC AI was eliminated in the overall assessment, and Naver Cloud was judged not to have met the independence criteria because it partially used a model that had already been trained by another company. After an additional call for applications, Motif joined, bringing the total to four teams for the second evaluation.

This time, all four teams developed models using their own technology and thus met the independence criteria. The evaluation scoring consisted of 40 points for benchmarks, 35 for expert evaluations, and 25 for user evaluations, for a total of 100 points. Benchmarks, which test AI performance in mathematics, coding, reasoning, and other areas, included 25 points from the “Intelligence Index (AAII)” of global AI evaluation agency Artificial Analysis (AA), and 15 points from the evaluation by the National Information Society Agency (NIA), which assesses Korean-language contextual performance as well as safety and reliability.

However, the MSIT did not disclose the total composite scores or the detailed scores by team. As a result, the only figure made public was AAII, which accounts for 25% of the total score. Motif’s “Motif 3,” which was eliminated in this round, recently scored 47 points in AAII, ranking first among top-tier Korean teams and 10th worldwide among flagship models by AI company. Second Vice Minister of Science and ICT Ryu Je-myeong stated, “Motif’s technological capability is at a world-class level,” but explained the reason for its elimination by adding, “Other models were ahead in usability and practical applicability.”

Within the AI industry, some say Motif’s elimination was predictable. Motif entered the competition belatedly through the additional call in February this year, after the first evaluation, so its model development and GPU support period were shorter than those of the three existing teams that had received support from the outset of the project. Nevertheless, it is being assessed as having proven its technological competitiveness by delivering performance in the global top 10. Motif CEO Lim Jeong-hwan said, “Regardless of the outcome, this has demonstrated that it is possible to develop AI models at the global frontier level based solely on technological capabilities in Korea,” adding, “We will continue with model development going forward.”

The three teams that passed this stage showed strengths in terms of real-world application. Upstage received high marks for integration with domestically produced AI semiconductors, while SK Telecom scored well for industrial applications in sectors such as defense and manufacturing. LG AI Research Institute was positively evaluated for improving model reliability, including reducing “hallucinations,” where AI generates incorrect answers.

● Passing teams to receive 1,000 GPUs each

The government will provide each of the three successful teams with 1,000 NVIDIA B200 GPUs in the second half of this year (July–December). Converted to a six-month rental cost, this amounts to about KRW 40 billion per team, or a total of KRW 120 billion. The third evaluation will proceed as originally planned, with two teams to be finally selected from among the three early next year.

However, the project structure may change after the final two teams are chosen. As the development pace of top-tier global AI accelerates, the government is reviewing ways to expand the scale of support. How to link this Dokpamo selection with the artificial general intelligence (AGI) development project is also expected to be decided soon.

Popular News