[Focus] Three-Player Survival in the Independent AI Foundation Model Contest… Three Distinct AI Utilization Strategies

In the second evaluation of the government’s Independent AI Foundation Model project, the surviving teams were SK Telecom (SKT), LG AI Research, and Upstage. As a result, the competitive field was narrowed to three organizations.

Motif Technologies, which joined through an additional February open call, was eliminated after about 6 months. The project’s open call began in June 2025, and the project period is longer than 1 year.

The remaining schedule consists of a third evaluation. The government plans to select the final two teams in early next year through that third evaluation.

The key evaluation point in the remaining competition is not only AI model performance, but also how the models are used. This article focuses on reviewing how the three organizations still in the race are using the development models linked to the Independent AI Foundation Model project.

Among them, SKT’s utilization model is 'A.X K2.' SKT is pushing ahead with applications in industrial settings, and the target cases include defense, manufacturing, automobiles, and gaming.

Defense is the field where its achievements stand out. In May, SKT exchanged an MOU with the Ministry of National Defense on 'using the Independent AI Foundation Model in the defense sector,' with the aim of accelerating AX in defense. SKT and the ministry will push ahead with the development and demonstration of defense-specific AI models based on the Independent AI Foundation Model, and they agreed to cooperate on collecting and providing open defense data.

SKT provided the lightweight A.X K1 model to the Ministry of National Defense. The lightweight version reduces size and improves processing speed. SKT applied additional training using defense data to K1 and K2, with the aim of optimizing the models. It also developed three Quantization models based on A.X K1 in cooperation with the ministry.

The manufacturing collaboration targets are KG Steel and Connec. KG Steel is a steelmaker, and Connec is an auto parts manufacturer. The purpose of the collaboration is to verify whether the independent AI model can be applied in the field. Since April, SKT has used data held by KG Steel and Connec, including past process error and accident analysis reports, equipment manuals, and logs. Based on that, SKT developed a demo version of a 'manufacturing-specific AI agent' based on A.X K1.

SKT will conduct production-process demonstrations in the second half of the year. SKT and KG Steel will apply the demo version of the manufacturing-specific AI agent to the cold-rolling line at KG Steel’s Dangjin plant, and SKT and Connec will apply the demo version to Connec’s casting and processing operations. Through this, they plan to verify whether AI can be used for process analysis and problem response based on accumulated manufacturing data.

SKT also applies A.X to its consumer services. The lightweight A.X K2 model is used in the AI calling agent 'Adot,' and part of the lightweight A.X K2 model is also used in a feature that organizes the content recorded in Adot Note by purpose.

LG AI Research is focusing on using vertical models. While continuing to build additional industry-specific AI models based on the foundation model 'EXAONE,' LG AI Research has proposed a direction in which specialized models handle problem response in industrial settings such as manufacturing and finance.

At the same time, 'K-EXAONE,' developed through the Independent AI Foundation Model project, is being deployed in competition with global general-purpose models. LG AI Research’s strategy is a two-track approach that runs global general-purpose models alongside industry-specific specialized models.

LG AI Research and LG Household & Health Care used 'EXAONE Discovery,' an AI model specialized in discovering new materials, to successfully develop a new cosmetic efficacy ingredient. 'EXAONE Discovery' analyzes large amounts of material molecular data and predicts properties based on the results, reducing the time and cost of finding candidate substances for research use. According to LG Household & Health Care, candidate selection used to take about 2 years, but with 'EXAONE Discovery,' it was completed in just one day.

In finance, 'EXAONE Business Intelligence' was deployed. In April, LG AI Research exchanged a memorandum of understanding with London Stock Exchange Group (LSEG) and Kiwoom Securities, and the core of the memorandum was a financial AI agent service. Accordingly, the three parties are pushing ahead with cooperation, refining the predictive score calculation function based on 'EXAONE Business Intelligence' and beginning development of a question-answering financial AI agent.

One of LG AI Research’s strengths is that it is expanding AI infrastructure and service deployment together, grounded in the group’s ecosystem. It is preparing an AI data center (AIDC) with LG Uplus and Fasoo, and the target completion date for the Fasoo AI data center (AIDC) is 2027. This project is centered on concentrating the capabilities of LG affiliates such as LG Electronics and LG Energy Solution, and the planned AIDC scale is capacity for up to 120,000 GPUs. Earlier, LG Uplus had already installed EXAONE in the lifestyle-focused AI agent 'ixi-O.'

LG AI Research is also active in cooperation with external companies beyond the group. ESTsecurity worked with LG AI Research to unveil 'Alyac Agentic EDR,' which is composed of a conversational chatbot and AI agents that assist across the security pipeline. The model applied there is K-EXAONE developed by LG AI Research, marking the first case of an EDR solution using the Independent AI Foundation Model.

Upstage is developing its business with a focus on supporting On-premise environments and targeting the public sector, and it is also pushing ahead with the next AI transition. In the legal and administrative fields, where there is a lot of text and document processing, use of 'Solar' is especially prominent.

One representative example is the Public AI Workspace supplied to the Public Procurement Service. Upstage signed a contract with the Public Procurement Service at the end of last year to supply a 'generative AI work support service' for public-sector tasks, and it was the first supplier of that service. As a result, it was listed on the Public AI Workspace in the Public Procurement Service’s Digital Service Mall.

The 'Public AI Workspace' was implemented by integrating Upstage's 'Solar' with document-processing AI technology based on optical character recognition (OCR), and it was built in agent form. It also supports various formats such as hwp files, Word, and PDF.

In the same context, Upstage’s model is also being used in deliberations on research and development (R&D) budgets. Starting this year, the Ministry of Science and ICT has been using a 'budget review-specialized AI' for the allocation and adjustment of national R&D budgets. Each year, 166 expert committee members and relevant staff participate in the budget review, and the government plans to improve efficiency by introducing AI into the process.

The budget review-specialized AI was developed based on Upstage’s 'Solar Open 100B,' with the Korea Institute of Science and Technology Information (KISTI) and the Electronics and Telecommunications Research Institute (ETRI) taking part in development. This AI used 5,000 data sets as training data, including R&D budget requests and planning reports accumulated over the past 5 years.

Meanwhile, after acquiring the operator of the Daum portal, 'AXZ,' in May, Upstage is pushing ahead with a portal transformation that incorporates AI technology into Daum. The portal site Daum (Daum) is expected to become Upstage’s public gateway for Solar aimed at general users.

The first example of this change is 'AI Summary.' Daum unveiled a search-result summary feature in December last year, and the model used at the time was not specified. After acquiring Daum, Upstage applied a single Solar model throughout the AI Summary generation process, and by the end of the year it plans to provide 'AI Mode,' a search-result-based chatbot and additional question-and-answer feature.

Source: IT DAILY · Kim Ho-jun, Seong Won-young
Original: https://www.itdaily.kr/news/articleView.html?idxno=241060

References

• https://www.itdaily.kr/news/articleView.html?idxno=241060