Korea Advances Three Teams in Sovereign AI Project
South Korea's Ministry of Science and ICT advanced teams led by LG AI Research, SK Telecom, and Upstage from the second evaluation of its government-backed independent AI foundation model project on August 18. Motif Technologies was eliminated by a narrow margin after benchmark, expert, and user evaluations, the ministry said. The three remaining teams are set to receive expanded B200 GPU support before the field is reduced to two finalists by year-end.
South Korea's Ministry of Science and ICT advanced LG AI Research, SK Telecom, and Upstage to the next stage of its independent AI foundation model project on August 18, eliminating Motif Technologies from a four-team second-round evaluation.
The state-backed program, also known as the independent AI foundation model project, is intended to support domestic foundation-model development. According to the Ministry of Science and ICT, reported by ChosunBiz and The Chosun Daily, the surviving teams are LG AI Research's K-EXAONE 2.0, SK Telecom's A.X K2, and Upstage's Solar Open 2; Motif competed with Motif-3.
Evaluation favored more than raw model scores
The ministry assessed teams through benchmark tests worth 40 points, expert review worth 35 points, and user evaluation worth 25 points, according to The Chosun Daily, Asiae, and the Korea Herald. The Korea Herald reported that benchmark scores averaged 22.5 across the four teams, with four points between the first- and fourth-ranked entries. Expert-review scores averaged 28.8, with a 2.4-point spread.
User assessments produced the largest separation, according to the Korea Herald: the average was 17.6 points and the gap between first and fourth place was five points. The Chosun Daily reported that the ministry attributed Motif's elimination to relatively lower usability and applicability scores, despite what it described as strong technical capabilities.
The benchmark component incorporated the Artificial Analysis Intelligence and Performance Index and National Information Society Agency assessment results, The Chosun Daily reported. The ministry and National IT Industry Promotion Agency said all four teams scored above 31 on the AAII, exceeding major European and Canadian models in that measure. The same report said Korean models appeared alongside US and Chinese models in the score range above 40.
These results show that benchmark performance can be necessary for progression, but evidence of usability, real-world task performance, and industrial applicability can materially affect selection outcomes.
More B200 compute, then another cut
The Ministry of Science and ICT said the three remaining teams will receive expanded access to B200 GPU-equivalent computing resources. According to ChosunBiz and the Korea Herald, support is set to increase from the equivalent of roughly 768 B200 GPUs in the first half of 2026 to about 1,000 in the second half.
The Korea Herald reported that additional evaluations will reduce the three teams to two finalists by the end of the year. The ministry described the added compute as support for developing more competitive domestic models.
Asiae reported that the expert evaluation credited Upstage's proposed integrations with platforms including Daum and Timely, and its collaboration with FuriosaAI on neural processing units. It reported that SK Telecom received strong marks for mathematical reasoning, Korean-language performance, and deployment in large-scale commercial services. Asiae also reported that LG AI Research was recognized for its approach to agentic AI and international collaboration efforts.
For ML practitioners, the selection criteria are notable because they combine standardized model assessment with expert and user-facing evidence.
Key Points
- 1South Korea retained LG AI Research, SK Telecom, and Upstage after a multi-part evaluation of domestic foundation-model contenders.
- 2Motif's elimination shows that usability and industrial applicability can outweigh strong technical performance in public model-selection programs.
- 3Expanded B200 GPU support gives the remaining teams more compute for model development before South Korea selects two teams by year-end.
Scoring Rationale
The program is a notable national effort to build competitive domestic foundation models and allocates substantial B200-equivalent compute to the remaining teams. Its evaluation framework is relevant to practitioners building production models because it weights usability and application readiness alongside benchmarks.
Sources
Public references used for this report.
View 1 more source
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
