The Ministry of Science and ICT and the National IT Industry Promotion Agency (NIPA) released the results of the second-stage evaluation of the ‘Sovereign AI Foundation Model’ project on the 18th. All four elite teams that took part demonstrated world-class performance, and the teams from Upstage, SK Telecom and LG AI Research (in alphabetical order) will advance to the next stage.
 [코리안투데이] Results of the second-stage evaluation of the ‘Sovereign AI Foundation Model’ project © 장형임 기자 |
The Ministry of Science and ICT (led by Minister and Deputy Prime Minister Bae Kyung-hoon) and the National IT Industry Promotion Agency (President Park Yoon-kyu) announced the results of the second-stage evaluation of the ‘Sovereign AI Foundation Model’ project on the 18th. Part of the national agenda to ‘build the country that uses AI best in the world,’ the project aims to secure large-scale AI models built with domestic technology, overcoming the resource gap with global big tech.
All four elite teams participating in the second stage released the AI models they developed in the first half of this year as open source. Motif Technologies unveiled ‘Motif 3 (314B-A13B),’ an agent-specialized model that recorded the world’s best performance (by AAII) among AI models from countries other than the United States and China. Upstage presented ‘Solar Open2 (250B-A15B),’ a cost-efficient model that can process hundreds of pages of documents at once and runs on just two NVIDIA H200 chips. SK Telecom released ‘A.X K2 (688B-A33B),’ which posted top-level results on Korean-language and mathematics benchmarks, while LG AI Research unveiled ‘K-EXAONE 2.0 (750B-A37B),’ the country’s largest model at 750 billion parameters, supporting 10 languages.
All four models were listed among the ‘Notable AI Models’ of Epoch AI, a U.S. nonprofit AI research institute. As of this year, only three countries — the United States, China and Korea — have models on this list. On the AAII, a composite index produced by the AI model evaluation platform Artificial Analysis, all four teams scored 31 points or higher, surpassing major models from Europe and Canada; Korean models joined those of the United States and China in the above-40-point tier.
The gap with top-tier models also narrowed. The difference of 19 points at the January evaluation (GPT-5.2 at 51 points versus the top domestic model at 32) shrank to 16 points this August (Claude Opus 5 at 63 points versus the top domestic model at 47). Among open-weight models — those that disclose their weights — the domestic elite teams’ models ranked 6th in the world on the AAII, and by each company’s top-scoring model, Motif Technologies’ ‘Motif 3’ entered the global top 10 with 47 points. In a comparison of each country’s best-performing language models, Korea also held third place after the United States and China.
Each model’s strengths stood out. Upstage’s ‘Solar Open2’ boasts a context window of up to one million tokens, giving it information-processing capacity on par with the world’s top-tier models. SK Telecom’s ‘A.X K2’ scored at the gold-medal threshold on the 2026 International Mathematical Olympiad (IMO) problem evaluation and achieved a tied-best 97.1% accuracy on the MathArena AIME 2026 leaderboard. LG AI Research’s ‘K-EXAONE 2.0’ ranked 9th in the world on a metric for minimizing hallucination, proving its reliability. Hugging Face, the world’s largest open-source platform, praised the project as an exemplary case of government-led open-source policy and, positioning Korea as an ‘alternative country’ beyond the United States and China, has proposed cooperation such as joint promotion — with concrete plans now under discussion.
Beyond the elite teams, various companies including KT, Kakao, Naver and NC AI are delivering results in industrial fields such as semiconductor processes, case-law analysis and new-material development, as well as in public sectors such as government administrative networks and R&D budget review. This shows, the government explained, that the project’s benefits are spreading throughout the domestic AI ecosystem.
The second evaluation was conducted using a combination of benchmark evaluation (40 points), expert evaluation (35 points) and user evaluation (25 points). The benchmark portion comprised the AAII benchmark (25 points) and the NIA benchmark (15 points), with measurements taken directly by Artificial Analysis, while 10 external experts with no conflicts of interest took part in the expert evaluation. The user evaluation was carried out as an absolute assessment by 49 professional AI users and 185 members of the general public — notably introducing, unlike the first round, a citizen-participation usability assessment. The overall averages for the four teams were 22.5 points for benchmarks, 28.8 for expert evaluation and 17.6 for user evaluation, and the score gap between first and fourth place was within a narrow 5 points in each category.
The elite teams advancing to the next stage will see their high-performance GPU (B200) support expanded from around 768 units in the first half of the year to about 1,000 units in the second half. The Ministry of Science and ICT said moves to treat AI models as national-security strategic assets are becoming visible among leading nations, and that it plans to discuss a new support framework for developing top-tier sovereign AI models with relevant ministries and announce the details later.
The results also carry significance for everyday life. As AI models built with domestic technology reach a level capable of competing with the world’s best, a foundation has been laid to reduce dependence on overseas services and to let the public directly experience safe and reliable AI services tailored to the Korean language and Korean context.
지역 홍보대사 신청 기자양성과정 신청 코리안투데이 지부장 신청 코리안투데이 인천강화
THE KOREAN TODAY × WIA CODE
이 기사의 WIA Code
WIA Code 스캐너로 비추면 인터넷 없이 발행 사실과 위·변조 여부를 확인할 수 있습니다.
휴대폰 기본 카메라로 찍으면 스캐너 안내 페이지가 열립니다 — 기사 바로 열기
움직이는 코드는 스트림입니다 — WIA Code 스캐너로 약 22초 비추면 기사 전문(사진 포함)이 폰에 저장되어 인터넷 없이도 읽을 수 있습니다.
WIA Code · 코리안투데이
Leave a Reply