LG AI Talk Concert 2025 (Image courtesy of Yonhap)
SEOUL, Jan. 15 (Korea Bizwire) — South Korea has narrowed the field in its push to develop a homegrown artificial intelligence foundation model, advancing three teams while cutting two others after a first-round evaluation that placed heavy weight on technological capability and the independence of model design.
The Ministry of Science and ICT said Thursday that LG AI Research, Upstage and SK Telecom passed the initial review of its “Independent AI Foundation Model Project.” NC AI and Naver Cloud were eliminated.
The evaluation, conducted with input from participating teams, combined benchmark testing, expert review and user assessments. LG AI Research emerged as the clear leader, topping all three categories.
It scored highest in benchmark tests measuring core capabilities such as reasoning, knowledge, long-form comprehension, reliability and safety, and also received the strongest marks from outside experts who reviewed technical reports and training logs to assess the completeness and originality of each model’s development process.
A visitor uses SK Telecom’s A.X K1 model at the first briefing of the Independent AI Foundation Model Project, held on Dec. 30 at COEX in Seoul’s Gangnam district. (Yonhap)
In user testing by AI professionals, LG again placed first, earning top scores for practical usability and cost efficiency.
The government declined to release detailed scores for individual companies beyond LG, saying it wanted to minimize potential reputational harm. Ryu Je-myeong, the ministry’s second vice minister, said the project was intended not to crown a single winner but to raise the global competitiveness of domestic AI developers through structured competition.
Naver Cloud’s elimination drew particular attention. The ministry concluded that the company failed to meet the project’s requirement for an “independent” AI model after it was disclosed that Naver Cloud had used encoder architecture and weights from Alibaba’s Qwen model.
Under the government’s definition, an independent foundation model must be designed and pretrained domestically, not derived from fine-tuning or reusing pretrained weights from overseas models.
“Initializing weights and training a model from scratch is the minimum condition for technical independence,” the ministry said, adding that using existing weights amounted to benefiting from another party’s training process.
While the government stressed that open-source software is not inherently problematic, officials said the purpose of the project was to ensure that Korean developers could design and train large-scale models on their own.
The LG AI Research booth at the first presentation of the “Sovereign AI Foundation Model” project, held on Dec. 30 at COEX in Gangnam, Seoul. (Yonhap)
Upstage and SK Telecom were flagged for minor ethical issues related to disclosures of reference materials, but the ministry said those concerns were not serious enough to affect their advancement.
The project reflects Seoul’s broader ambition to establish a sovereign AI ecosystem amid intensifying global competition, even as it seeks to balance openness to global innovation with clearer standards on technological self-reliance.
Kevin Lee (kevinlee@koreabizwire.com)
