Motif eliminated from national AI model project as LG AI Research, SK Telecom, Upstage advance

Motif has been eliminated from the second round of South Korea's national AI foundation model project, known as "Dokpamo," leaving LG AI Research, SK Telecom and Upstage to compete for the two final spots to be decided by year-end.

What separated the companies in the second round was their ability to demonstrate technology applicable in real-world settings. Rather than raw benchmark scores, "usability" proved decisive — and the government said the results will accelerate its push to secure sovereign AI capabilities.

The Ministry of Science and ICT announced the results Tuesday at the Government Seoul Building in Jongno-gu, Seoul, confirming that LG AI Research, SK Telecom and Upstage had advanced to the final evaluation while Motif was eliminated.

The centerpiece of the second-round evaluation was demonstrable, real-world usability. Scoring was divided across four categories: the AI Intelligence Index (AAII) worth 25 points, a National Information Society Agency (NIA) benchmark worth 15 points, expert evaluation worth 35 points, and user evaluation worth 25 points.

The assessment went beyond expert and benchmark reviews to include a public evaluation in which 200 ordinary citizens directly tested each team's AI model from Aug. 8 through Aug. 11 and scored them on an absolute scale.

A high benchmark score alone was not enough to secure a spot as the national AI representative. In the AAII metric, Motif3 scored 47 points, Upstage 37, SK Telecom 35 and LG AI Research 31 — but those figures were not the deciding factor.

Each company poured its efforts into improving usability. LG AI Research's K-Exaone was recently integrated into the portal Zum's AI search feature, with Zum replacing the core AI models powering its "AI 3-second summary" and "AI issue trends" services with K-Exaone.

SK Telecom's A.X K2 improved average performance across 14 domestic and international benchmarks by 32.2 percentage points compared with its predecessor, A.X K1, with particularly strong gains of about 83.9 percentage points in long-context comprehension and agent-related evaluations.

Upstage, building on its Solar WBL model, moved beyond deploying Solar on the Daum portal it acquired and signaled plans to expand into sectors including finance and manufacturing. Motif Technologies, meanwhile, disclosed interim evaluation results for Motif3 — a model with about 314 billion parameters — and matched the score of DeepSeek's latest model, V4 Pro, on the AAII metric with 44 points.

With the second-round results now announced, the three remaining teams will enter a final competition for two spots to be decided by year-end.

The three teams that advanced are expected to receive not only computing resources such as GPUs but also government research and development support.

Advancing teams also gain a symbolic edge as the "national AI representative" — a designation that carries weight in the broader ecosystem-building race, beyond the government's separately run "AI for All" initiative, a multimodal and multi-agent platform whose public application period closed Tuesday. Industry observers expect that companies able to translate benchmark performance into real service competitiveness — and secure GPUs, talent and ecosystem partnerships — will emerge as the dominant players in South Korea's AI landscape.

The government expressed strong confidence in the technical capabilities of the domestic AI models, saying the second-round results put South Korea within reach of moving beyond third place globally to challenge for second.

Deputy Prime Minister and Minister of Science and ICT Bae Kyung-hoon said at a presidential briefing held at Cheong Wa Dae last month that he expected the August second-round results to show South Korea could "go beyond the existing third-place ranking and challenge for second place."

"Korea can build AI models at that level," Bae added. "If the government properly prepares the infrastructure and provides support, we can build models that match that standard."