LG AI Research, SK Telecom, Upstage survive next cut in national AI model project
Motif cut despite strong global benchmark results
Graphic by Cho Sang-won
The government has narrowed its state-backed project to develop a domestic artificial intelligence (AI) foundation model to three teams led by LG AI Research, SK Telecom and Upstage, eliminating startup Motif Technologies despite its strong showing on global benchmarks.
The Ministry of Science and ICT announced Tuesday that the three consortia advanced after the second evaluation of four contenders, which assessed benchmark performance, expert reviews and user experience.
The project, launched last year to develop homegrown AI models that can rival leading United States and Chinese models, initially selected five teams for the first stage. But after two teams led by Naver Cloud and NC AI were eliminated in the first evaluation in January, Motif joined the competition through an additional selection process to fill the vacant spot.
The four teams submitted their latest models for the second-round assessment. Their models included LG AI Research’s K-EXAONE 2.0, SK Telecom’s A.X K2, Upstage’s Solar Open2 and Motif’s Motif 3.
The ministry said all four teams had developed internationally recognized models, with all four recognized by U.S. nonprofit AI research organization Epoch AI as notable AI models and scoring above 31 on the Artificial Analysis Intelligence Index (AAII), a global benchmark for AI model performance.
"Above all, compared with the first stage, the performance gap between the nominated teams' models in the second stage and the world's top AI models has narrowed, further strengthening Korea's position as one of the world's three leading AI powers after the U.S. and China,” Second Vice Science Minister Ryu Je-myung said during a briefing at Government Complex Seoul on the day.
The second evaluation allocated 40 points to benchmarks, 35 points to expert assessments and 25 points to user evaluations. The four teams scored an average of 22.5 points in benchmarks, 28.8 points in expert assessments and 17.6 points in user evaluations.
Although the ministry did not disclose individual rankings, the gaps between the highest- and lowest-scoring teams were 4 points in the benchmark assessment, 2.4 points in the expert assessment and 5 points in the user evaluation, with a different team ranking first in each category.
Second Vice Science Minister Ryu Je-myung speaks during a briefing at Government Complex Seoul, Tuesday. Yonhap
The ministry said Motif 3 scored 47 on the AAII, ranking 10th worldwide and making it the highest-performing model from a country outside the U.S. and China.
However, its benchmark strength was not enough to secure advancement as the evaluation placed greater weight on practical use and ecosystem impact.
"Motif showed outstanding technological capabilities, but received a somewhat lower evaluation than other companies in terms of usability and applicability, areas that carried a substantial weight of the 75 points in this evaluation," Ryu said.
The ministry said LG AI Research’s model was highly rated for reliability and risk management, while SK Telecom’s model displayed a top-tier performance in math and Korean and demonstrated strong usability through large-scale commercial deployments. Upstage’s model stood out for its integration with internet portal Daum and collaboration with FuriosaAI on domestic AI hardware.
The government plans to boost graphics processing unit (GPU) support for the three teams advancing to the next round. It will provide about 1,000 Nvidia B200 GPUs in the second half of this year, up from around 768 in the first half, to support further development of their models.
The third-stage evaluation will proceed in the second half of the year, but its detailed scoring system has not yet been finalized. The ministry said it will first complete an objection period for the second-stage results and then consult with the three remaining teams on the evaluation criteria and procedures.
Separately, the ministry is reviewing plans to support the development of frontier-level AI models as global competition intensifies. It will also consider how the initiative could be linked to or reshape existing programs, including the national foundation model project and the AGI project.
The ministry said it plans to announce its broader direction for the frontier-level AI model initiative soon, including whether to continue the national foundation model project beyond its third stage.