South Korea’s Motif Technologies—which ranked 10th globally on a key AI benchmark—has been eliminated from the government’s “Sovereign AI Foundation Model” project, while three teams—Upstage, SK Telecom, and LG AI Research—have advanced to the next phase.
South Korea’s Ministry of Science and ICT (MSIT) and the National IT Industry Promotion Agency (NIPA) announced on the 18th that they had finalized the results of the project’s second-phase evaluation. The assessment covered four models: LG AI Research’s “K-ExaOne 2.0,” SK Telecom’s “A.X K2,” Upstage’s “Solar Open 2,” and Motif Technologies’ “Motif 3.”
The evaluation consisted of benchmark testing (40 points), expert assessment (35 points), and user evaluation (25 points). In benchmark testing, the four teams averaged 22.5 points, with a gap of just 4.0 points between first and fourth place. Expert assessment averaged 28.8 points with a 2.4-point spread, while user evaluation averaged 17.6 points with a 5.0-point gap. “In the composite evaluation, Motif Technologies was not selected due to a narrow margin,” MSIT explained.
Motif Technologies’ “Motif 3” scored 47 points on the Artificial Analysis Intelligence Index (AAII), a global AI evaluation platform, ranking 10th among AI models worldwide and the highest among the four participating models in the project. The model—developed independently from architecture design through implementation with 314 billion parameters—earned recognition for its technical prowess, but received lower marks than other teams in usability and practicality.
Ryu Je-myung, Second Vice Minister of Science and ICT, said at a briefing that day: “Motif clearly demonstrated outstanding performance with an exceptionally high AAII score. However, out of a total of 100 points, AAII accounts for only 25 points. When combining other evaluations, while the technology was excellent, the remaining 75 points—heavily weighted toward usability and practicality—saw lower scores compared to other companies.”
Vice Minister Ryu drew a line regarding recent benchmarking controversies (excessive optimization for specific test sets), stating that such concerns “were not reflected in the evaluation.” He added: “Through AAII’s official response and expert assessment, we concluded there was no unfair technical support. We received confirmation that no evidence was found to substantiate systematic memorization or overfitting issues.”
Differentiated Strengths of the Three Teams
The advancing teams were recognized for strengths in distinct areas. SK Telecom’s A.X K2 secured top-tier performance in mathematical reasoning and Korean-language capabilities among the comparison group, recording a score equivalent to a gold medal standard on the 2026 International Mathematical Olympiad problem set. It was also evaluated as having demonstrated potential for application in large-scale commercial services.
LG AI Research’s K-ExaOne 2.0 delivered results in reliability metrics for reducing hallucinations. Among AAII benchmark indicators, it ranked 9th globally in AI model hallucination minimization and also showed strength in multilingual support.
Upstage’s Solar Open 2 leveraged its ability to process contexts of up to 1 million tokens—capable of handling documents spanning hundreds of pages in a single pass. The company plans to deploy it in AI agent services for complex tasks such as long-document analysis, information retrieval, and report generation. Notably, the model’s integration with FuriosaAI’s domestically produced neural processing units (NPUs) to lower service operating costs received high marks.
Solar Open 2 is designed with a “250B-A15B” architecture that activates 15 billion of its total 250 billion parameters during inference. It can run on just two Nvidia H200 GPUs, offering an advantage for companies without large-scale GPU infrastructure. Upstage plans to integrate the model with the “Daum” portal and “Timely” platform to validate response performance, stability, and cost efficiency in real-world usage environments.
Third-Phase Evaluation and Expanded GPU Support
The government plans to expand Nvidia B200 GPU support for the three teams advancing to the third phase from approximately 768 units in the first half of this year to roughly 1,000 units in the second half. Choi Dong-won, a director at MSIT, explained: “Assuming a six-month GPU rental period, we estimate the cost at around 40 billion won (approximately $28.4 million). Converting the GPUs to be provided to the three teams into monetary terms, we expect roughly 120 billion won (approximately $85.1 million).”
The project originally envisioned selecting two final companies through a third-phase evaluation, with these companies pursuing AI model development at approximately 95% of global-leading capability by 2027. However, Vice Minister Ryu emphasized: “This project is not about picking one or two winners. The goal is to provide direct and indirect stimulus for South Korea’s AI ecosystem to develop to global standards through competition.”
The government is also reviewing a restructuring of the competitive framework to respond to the exponential advancement of global frontier companies’ model performance. “We are engaged in extensive discussions with the companies themselves—the actual stakeholders—on whether to distribute or concentrate resources, and what kind of consortium structures would be desirable,” Ryu said.
The specific evaluation methodology and scoring weights for the third phase have not yet been finalized. MSIT plans to review supplementary items identified in the first and second evaluations and consult with the three elite teams to finalize and announce the evaluation plan. An appeals process remains open; if objections are filed, the process will proceed after a clarification procedure.
Meanwhile, the second-phase user evaluation involved 49 AI expert users and 185 members of the general public. The general public evaluation panel originally targeted 200 participants, with approximately 1,400 applicants. MSIT stated that public evaluations did not influence the final selection outcome.