Korea's national AI race enters decisive month as second evaluation looms

A file image related to the Independent AI Foundation Model Project [Generated by AI]

The decisive month has arrived for the Independent AI Foundation Model Project, the government initiative to select Korea's national AI.

All four elite teams — LG Group's AI Research Institute, SK Telecom, Upstage and Motif Technologies, which joined through a supplementary open call — have completed development of their proprietary models. A public evaluation running from Saturday through Tuesday will feed into the second-round results, which will narrow the field from four teams to three and are expected to be announced shortly.

The government has expressed confidence that the second evaluation will demonstrate technology capable of challenging not just a top-three global ranking but a top-two position, raising expectations that Korea's push for sovereign AI is gaining real momentum.

According to industry sources, the second-round results — cutting the four elite teams to three — could be revealed as early as Monday.

Motif Technologies is set to submit its final model, Motif 3, along with a performance report to the Ministry of Science and ICT and the National IT Industry Promotion Agency on Tuesday. With that submission, all four elite teams will have completed development of their independent AI models, following LG Group's AI Research Institute, SK Telecom and Upstage, which submitted their models in early July.

The second evaluation goes beyond expert and benchmark assessments to include a public evaluation in which ordinary citizens test the four teams' AI models directly. The government is recruiting 200 members of the public to serve as evaluators, selected using gender and age quotas. The panel will use each team's AI model from Saturday through Tuesday and score them on an absolute scale.

Given the government's stated commitment to accelerating the project, results are likely to be announced quickly once the public evaluation concludes. Growing expectation holds that an outline of the outcome could emerge as early as Monday.

The final selection — narrowing three teams to two — is scheduled to be completed within the year. The government had originally planned to name the final two teams in February next year, but has since decided to finish the entire project schedule before year's end.

Visitors gather at the A.X K1 demonstration booth during the first public presentation of the Independent AI Foundation Model Project on Dec. 30 last year. [Provided by SK Telecom]

The focus of the second evaluation is applied technology — AI that can be deployed in real industrial settings. The broad framework from the first round is being retained, with benchmark evaluation, expert evaluation and user evaluation all carrying over, but practical usability in actual industry environments is expected to be a key additional criterion.

Competition among the four elite teams has intensified in the final stretch. Ahead of the second evaluation, SK Telecom unveiled its large-scale AI model A.X K2, which has about 688 billion parameters. The company significantly expanded the model's scale compared with its predecessor and focused on deploying it in real industrial settings to build a Korea-centered AI ecosystem.

According to SK Telecom, A.X K2 achieved an average performance improvement of 32.2 percentage points across 14 domestic and international benchmarks compared with A.X K1. In long-context comprehension and agent-related evaluations, performance improved by about 83.9 percentage points.

LG Group's AI Research Institute has been expanding real-world deployments of its K-Exaone model, including a recent integration into the AI search feature of the portal Zum. Zum replaced the core AI models powering its "AI 3-Second Summary" and "AI Issue Trends" services with K-Exaone.

K-Exaone has already been deployed in the production and manufacturing processes of LG Group's manufacturing affiliates, broadening its ecosystem footprint.

Upstage is pursuing a strategy of targeting the finance and manufacturing sectors with its Solar WBL model. The company is also rolling out Solar on the portal Daum, which it acquired, positioning the model as an AI for the general public.

Motif Technologies released interim evaluation results for Motif 3, its foundation model with about 314 billion parameters.

According to Motif Technologies, Motif 3 scored 44 points on the AAII comprehensive performance index of Artificial Analysis, a global AI evaluation firm — matching the score of DeepSeek's latest model, V4 Pro.

The government has expressed strong confidence in the technical capabilities of Korea's homegrown AI models. Officials said they expect the second evaluation to demonstrate technology that can close the gap not just with the world's top three but with the top two.

Bae Kyung-hoon, deputy prime minister and minister of science and ICT, said at a presidential briefing held at Cheong Wa Dae last month that he expects the August second-round AI model evaluation results to show Korea can challenge not just a third-place ranking but a top-two position globally.

sjpark@heraldcorp.com