Government: Upstage, SKT, and LG Advance in Second Dokpamo Evaluation... Motif Misses Out Despite Benchmark Surprise

"1,000 Nvidia B200 GPUs to Be Provided to Three Teams"

Ryu Jemyung, Deputy Minister of the Ministry of Science and ICT, is announcing the results of the second evaluation of the independent AI foundation model project at the Government Seoul Office in Jongno-gu, Seoul on August 18, 2026. Photo by Yongjun Cho

In the second phase evaluation of the Ministry of Science and ICT's independent artificial intelligence (AI) foundation model (DOKPAMO) project, Upstage, SK Telecom, and LG AI Research Institute (listed alphabetically) have been selected as the final three models to advance to the next phase. Motif Technologies (Motif), which joined the evaluation later, recently caused a stir by ranking first in the Artificial Analysis Intelligence Index (AAII)—a comprehensive indicator by the AI performance assessment agency Artificial Analysis—but ultimately did not join the final models.

On August 18, the Ministry of Science and ICT held a briefing at the Government Seoul Office, announcing that three teams—SK Telecom, LG AI Research Institute, and Upstage—out of four participating companies, have survived and will advance to the next stage. The ministry stated, "In the second evaluation, the results of the benchmark evaluation (40 points), expert evaluation (35 points), and user evaluation (25 points) were aggregated, and Motif was eliminated by a narrow margin."

According to the expert committee’s evaluation, Upstage was praised for making its development achievements tangible for the public by promoting integration with the ‘Daum’ and ‘Timely’ platforms, as well as for alleviating reliance on foreign hardware via collaboration with FuriosaAI's neural processing unit (NPU). SK Telecom secured top-tier performance in both mathematical reasoning and Korean language tasks and received high marks for usability and practicality as it is already adopted in large-scale commercial services. LG AI Research Institute was recognized for establishing a collaboration strategy with global international organizations to increase global influence, as well as for distinguishing itself in agentic AI.

Motif, top benchmark scorer, excluded due to low usability evaluation

Motif's 'Motif 3,' presented in the second evaluation phase, scored highly in the benchmark evaluation but was ultimately eliminated in the comprehensive assessment, meaning Motif exited the race about six months after being newly selected for DOKPAMO. Motif joined the project in February. Despite having less preparation time than the competing consortiums, Motif developed its large language model (LLM), Motif 3, 'from scratch.' The Motif 3 boasts 314 billion parameters and was independently designed and built, adopting a Mixture of Experts (MoE) structure so that only 13.2 billion parameters actually function during operation, thus improving efficiency. Motif explains that even with fewer total and active parameters than comparable open-source models from China, Motif 3 delivered similar performance.

Motif also recorded the highest score among all participants in a global assessment agency’s benchmark for the DOKPAMO second evaluation. According to the AI performance evaluation agency Artificial Analysis, Motif 3 scored 47 points on the 'Artificial Analysis Intelligence Index (AAII),' outperforming Solar Open 2 (37 points, Upstage), A.DotX-K2 (35 points, SK Telecom), and K-XAOne 2.0 (31 points, LG AI Research Institute).

However, the benchmark score did not fully reflect actual model performance. In this second DOKPAMO evaluation, the benchmark was worth 40 out of 100 possible points. Of those, Motif's high AAII score accounted for 25 points, while the benchmark by the National Information Society Agency (NIA) made up 15 points. Expert evaluation (35 points) and user evaluation (25 points) comprised 60% of the assessment.

The limitations of the AAII benchmark also played a role. The AAII aggregates several sub-benchmarks across fields such as knowledge, reasoning, and coding into a single composite score, so grades can fluctuate significantly depending on which benchmarks are included. As a result, the NIA benchmark—another key assessment factor—evaluates overall performance, safety, and reliability in the context of Korean language and culture, enabling a more comprehensive assessment of domestic AI model competitiveness that cannot be determined by benchmarks alone.

Ryu Jemyung, Vice Minister of the Ministry of Science and ICT, is announcing the results of the second evaluation of the independent AI foundation model project at the Government Seoul Office building in Jongno-gu, Seoul. 2026.8.18 Photo by Cho Yongjun

Ministry of Science and ICT: "Final phase evaluation for three consortiums"

The ministry did not disclose the order of benchmark scores by model in the second phase evaluation. The average score for the four elite teams in the benchmark evaluation was 22.5 points, and the gap between first and fourth place was only 4.0 points. Vice Minister Ryu Jemyung of the Ministry of Science and ICT also mentioned that the user evaluation did not determine the final outcome in this round.

The government explained that while Motif was eliminated in this evaluation, it achieved results on par with global big tech companies in advanced benchmark fields such as agentic AI and coding. Vice Minister Ryu commented, "Motif demonstrated outstanding achievements despite being a small-scale R&D organization," and noted, "The reason for Motif's elimination in the second phase appears to be its relatively lower usability score compared to others."

The government plans to expand high-performance GPU (B200) support for the three teams advancing to the next stage by increasing the number of GPUs supported from 768 to around 1,000. Vice Minister Ryu added, "We deeply recognize the necessity to develop top-tier independent AI models and are discussing new support systems that go beyond existing approaches with relevant ministries. We will prepare and announce more specific measures going forward."

This content was produced with the assistance of AI translation services.