[E-Daily Reporter Kim Hyun-ah ] An unexpected outcome emerged in the second round of evaluations for the Ministry of Science and ICT’s Independent AI Foundation Model Project (DOKPAMO).

While Motif Technologies—which had ranked first in the AAII, the comprehensive AI model evaluation conducted by the global AI evaluation agency Artificial Analysis—was eliminated, the LG Corp.(003550) AI Research Institute, which had the lowest AAII score, advanced to the third round of evaluation. Upstage and SKTelecom(017670)also passed the second round.

These results do not simply mean that Motif lagged in performance while LG Corp. took the lead. They are interpreted as a signal that the competition criteria are shifting toward a comprehensive evaluation of technological capabilities, practical applicability, and the ability to expand the ecosystem—factors necessary for growth into a global frontier AI—beyond the benchmark performance of current models.

Motif, which ranked first in the AAII, was eliminated in the overall evaluation

Motif 3 scored 47 points in the AAII, significantly outperforming Upstage (37 points), SKTelecom (35 points), and LG AI Research (31 points). When converted to the government evaluation’s AAII weighting of 25 points, Motif scored 11.75 points, Upstage 9.25 points, SKTelecom 8.75 points, and LG Corp. 7.75 points.

Motif 3 is a MoE (Mix of Experts) model with a total of 314 billion (314B) parameters, of which approximately 13.2 billion (13.2B) are activated during the actual inference process.

However, the overall evaluation was different. The second-round evaluation was conducted on a 100-point scale, with 40 points for global benchmarks, 35 points for expert evaluations, and 25 points for user evaluations. Of these, AAII accounted for 25 points.

Ryu Je-myeong, Second Vice Minister of Science and ICT, explained, “Although the technical capabilities were outstanding, the AAII received a score of 25 points,” adding, “In the usability and applicability categories—which accounted for 75 points and carried significant weight—the company received a lower evaluation compared to other firms.”

The government did not disclose the detailed scores for each company and stated that no single item was the decisive factor in the rejection. It is interpreted that Motif was unable to sufficiently compensate in other evaluation areas for the high model performance it demonstrated in the AAII assessment.

LG Corp. 750B, SKT 688B… The Trend Toward Larger Models

Is

Also

Gaining Momentum

in Korea

Domestic AI models are also rapidly increasing

in scale

.

LG Corp. AI Research’s K-ExaOne has 750 billion (750B) parameters, while SKTelecom’s A.X K2 has 688 billion (688B) parameters.

Ahead of this evaluation, LG Corp. expanded the size of its model to more than three times its previous size. The strategy is to secure training and inference technologies for ultra-large models and lay the groundwork to compete in the same “weight class” as global frontier models.

Lim Woo-hyung, Co-Director of LG Corp. AI Research, stated, “It is difficult to narrow the technological gap with global Big Tech using only small models,” adding, “We will continue to take on this challenge with the goal of achieving performance equal to or better than that of models in the same scale category.”

However, the number of parameters alone does not equate to competitiveness. It must be supported by data, training efficiency, inference performance, agentic AI, practical service applicability, and safety.

Global models are in the 2–3 trillion-parameter range… We’ll start by securing 1 trillion parameters

The pace at which global AI models are scaling up is accelerating.

Among publicly released models, China’s Moonshot AI’s Kimi K3 is reported to have 2.8 trillion parameters (2.8T), Alibaba’s Qwen 3.8-Max has 2.4 trillion (2.4T), and DeepSeek V4 Pro is said to be around 1.6T.

In contrast, major closed-source frontier models such as OpenAI’s GPT, Google’s Gemini, and Anthropic’s Claude do not disclose their parameter sizes.

Therefore, it is difficult to conclude that all global frontier models are in the 10 trillion (10T) range. However, with the emergence of models exceeding 1 trillion and reaching the 2–3 trillion range, it is clear that parameters in the trillions have emerged as a key benchmark in the next-generation AI competition.

LG Corp.’s 750 billion (750B) parameters are massive by domestic standards, but there is still a gap when compared to the world’s top open models. This is why LG Corp. emphasizes “competing in the same weight class.”

From “No. 1 in Korea” to “Frontier AI”: The Future of Dokpamo

Amid this, the government has also hinted that the next phase of the Dokpamo project may differ from the existing approach.

Vice Minister Ryu stated, “Given that global frontier companies are advancing model performance exponentially, we have been engaging in ongoing discussions with companies regarding whether the current scale of support and competition format for independent foundation model projects can keep pace with these developments.”

He added, “There is a consensus that we should move beyond the existing competitive format and restructure it in a way that allows for competition with models on par with those of global frontier companies.”

The key point is not to focus on “the model that scores the highest domestically,” but to secure national-level AI capabilities that can continuously compete with the world’s top models.

Trillion-won Budget Investment… Possibility of Specific Companies Taking the Lead

Developing cutting-edge AI requires massive computing resources and capital. Deputy Prime Minister and Minister of Science and ICT Bae Kyung-hoon previously stated that developing a Mitos-class model would require approximately 10,000 of the latest GPUs, with the cost of the GPUs alone amounting to about 3.5 trillion won.

Accordingly, future government support is likely to expand to include not only fiscal investments in the trillions but also a model that combines private companies’ capital, technology, human resources, and data.

Kim Kyung-man, Director General of the Artificial Intelligence Policy Bureau at the Ministry of Science and ICT, commented on the Frontier AI project during a press briefing that day, stating, “Private-sector participation is necessary, and while forming a special-purpose company (SPC) involving multiple companies is one option, it is also possible for a specific company to lead the project.”

This implies that, moving beyond the “Dokpamo” model where multiple companies compete under identical conditions, a new model is possible in which companies with technological capabilities, infrastructure, and global collaboration capacity lead the development of frontier-level AI, while the government provides support for large-scale computing and R&D.

From “Who’s Number One?” to “Who Will Create the Next AI?”

From this perspective, the government plans to provide Upstage, SKTelecom, and LG Corp. AI Research—all of which passed the second round of the Dokpamo evaluation—with 40 billion won and 1,000 GPUs each to help improve model performance by the end of the year.

While the existing competition format—in which one company is eliminated at the end of the year—will be maintained, there is a possibility that the government will consider integrating the Dokpamo program with the “Frontier AI” initiative—which aims to develop models with parameters in the “jo” range and a scale of up to 10 trillion (10T)—once inter-ministerial budget negotiations for frontier-level model development are finalized.

Motif’s elimination and LG Corp.’s advancement in this round indicate what the next phase of competition in the Dokpamo program will entail.

Motif demonstrated overwhelming technical prowess in the AAII with its 314B (314 billion) model. LG Corp. advanced to the third evaluation by scaling up its model size to 750B (750 billion) and highlighting its comprehensive competitiveness in areas such as global collaboration, agentic AI, and safety. SKT has also joined the race for larger-scale models with a 688B (688 billion) model.

Going forward, it is increasingly likely that the competitive benchmark for the “Domestic AI Championship” will shift from “who received the highest score in this evaluation” to “who can sustainably develop global frontier AI.”

Amid discussions about government investments in the trillions of won and the potential for private-sector leadership, observers note that the “Domestic AI Model of the Year” competition has reached an inflection point, shifting from a contest to determine the “best domestic AI model” to a race to secure national-level AI capabilities capable of actually competing with the frontier AI systems of the United States and China.