South Korea's AI Race Heats Up as National Model Contest Nears Second Evaluation Cut
Translated from Korean, summarized and contextualized by DistantNews.
At a glance
- South Korea's government is nearing the announcement of the second evaluation results for its national AI foundation model project.
- Four teams, LG AI Research, Upstage, SK Telecom, and Motif Technology, are competing for three survival spots, showcasing diverse strategies.
- All four models have received recognition on the global stage, being listed by Epoch AI as 'notable AI models.'
South Korea's Ministry of Science and ICT is preparing to announce the results of the second evaluation for its "Foundation Model of Independent Artificial Intelligence" (Dokpamo) project. This critical stage narrows down the field of four competing teams, LG AI Research, Upstage, SK Telecom, and Motif Technology, to three survivors. The project aims to foster domestic AI capabilities, and the performance of these teams has already garnered international attention, with all four models being recognized by the U.S.-based Epoch AI as "notable AI models."
K-EXAONE 2.0
The second evaluation, which concluded recently, involved assessments from a 200-member public evaluation panel, benchmark tests, and expert reviews. The final decision, expected as early as next week, will eliminate one team from the competition. This rigorous process is designed to identify the most promising AI foundation models developed within South Korea.
A.X K2
Each of the remaining teams has adopted distinct strategies. LG AI Research has significantly scaled up its model, "K-EXAONE 2.0," boasting 750 billion parameters, more than triple the size of its previous iteration. SK Telecom's "A.X K2" model focuses on enhancing Korean language and math capabilities, as well as long-context processing, utilizing proprietary technology to improve efficiency. Upstage is championing a smaller, more efficient model, "Solar Open2," which, despite having the smallest parameter count at 250 billion, can handle up to 1 million tokens of context.
Solar Open2
Motif Technology, a late entrant to the second evaluation, emphasizes its proprietary design and computational efficiency. Their "Motif3" model utilizes a unique structure to reduce memory usage and focus on relevant information when processing long contexts. This approach yielded the highest score on the Artificial Intelligence Intelligence Index (AAII) from AI assessment firm Artificial Analysis. The Ministry plans to select two final teams through a third evaluation later this year, with the government providing support for hardware like GPUs and intending to utilize the developed models for public AI services.
Motif3
Originally published by Hankyoreh in Korean. Translated, summarized, and contextualized by our editorial team with added local perspective. Read our editorial standards.