South Korea Picks Five AI Teams to Challenge Global Models and Build Domestic Control

Five teams face a moving target in the global AI race

South Korea has selected five industrial consortia to build national artificial intelligence models, setting a target of at least 95 percent of the performance of the most advanced global systems released within the previous six months. Led by LG AI Research, Naver Cloud, SK Telecom, Upstage and NC AI, the teams will compete through repeated evaluations until two remain by 2027.

The Ministry of Science and ICT chose the five from 15 applicants for its Independent AI Foundation Models program. Each selected consortium carries the labels K-AI model and K-AI company. One team is scheduled to be eliminated in December, leaving four, with further reviews every six months.

As of September 7, 2025, the initiative brings together models already being developed for Korean search, business documents, customer service, coding and image understanding. LG has Exaone 4.0, Naver has HyperClova X, SK Telecom has its A.X series, Upstage has Solar Pro 2 and NC AI has Varco. Several companies say their systems outperform leading foreign models on tests of Korean language or cultural knowledge.

Seoul wants more than a domestic alternative to ChatGPT. Its sovereign AI strategy connects chips, data centers, cloud services and models, with the aim of keeping data, computing capacity and core expertise inside South Korea. A foundation model is a broadly trained AI system that can be adapted to many tasks. A large language model, or LLM, specializes in understanding and generating language.

The government is combining shared datasets, access to powerful computing hardware and encouragement for open source releases. Its challenge is to turn competitive test scores into reliable services while the global standard keeps advancing.

Why Seoul wants an AI system it can control

South Korean policymakers describe sovereign AI as an integrated national capability, rather than simply a chatbot made by a Korean company. They want systems that fit local laws, values and security needs, while reducing dependence on foreign providers for essential digital infrastructure.

Government leaders frame the effort as independence rather than isolation. South Korea could offer a Third Way between the United States and China, supplying capable technology to countries seeking partners without geopolitical strings attached. Export ambitions sit alongside the immediate goal of serving domestic businesses and public institutions.

ChatGPT's rapid rise convinced the Korean technology sector that it could not stay outside the foundation model race. China's DeepSeek then presented another challenge: its claims of much cheaper training at scale suggested that the economics of developing advanced models might change.

Seoul responded with flagship programs for an independent foundation model and a national AI computing center. These sit within a broader public and private investment agenda on the order of 100 trillion won. The government also raised the position of AI policy within the presidential office and moved to recruit researchers from abroad.

The investment agenda covers a wider undertaking than the five selected model teams. Its purpose is to give South Korea the ability to build, operate and export AI through a domestic network of hardware, infrastructure and software companies.

What the teams receive and when they face review

Each consortium receives access to processed datasets worth 10 billion won. The program also provides a pool of high quality broadcast and video learning data valued at 20 billion won, plus 2.8 billion won per team to develop datasets for particular fields.

SK Telecom and Naver Cloud will lease graphics processing unit infrastructure to support training and fine tuning for Upstage, NC AI and LG AI Research. GPUs perform many calculations at once, making them useful for training large models. Fine tuning means further training a model for a particular task or subject.

Upstage also receives matching support to recruit international researchers. Several teams say their first round of models will be released within the year, although specific release dates have not been given.

An ICT ministry official described the selection as a combination of technical capability and plans to make the technology usable by other businesses:

All five teams have demonstrated exceptional capabilities in AI model development, sharing a clear commitment to the vision of sovereign AI and presenting robust open source strategies that will allow other businesses to adopt and commercialize their technologies.

South Korea's ICT minister linked the competition to a broader national ecosystem:

Our bold initiative marks the beginning of Korea's journey toward building AI for all. The government will stand firmly behind AI companies and institutions as they scale new heights and shape a robust sovereign AI ecosystem.

The planned sequence gives the teams both an initial deadline and a longer competition:

• During 2025: several teams intend to release an initial round of models.
• December 2025: one of the five teams is scheduled to be cut, leaving four.
• Every six months: further evaluations will measure progress.
• By 2027: two teams are expected to remain.

LG and Upstage compete on reasoning and efficiency

LG AI Research is positioning Exaone 4.0 for enterprise work, including automating workflows, analyzing documents and turning unstructured text into organized information. For example, extracting structured information means finding useful details in prose and arranging them in a form that software can process.

Exaone has ranked near the frontier on global indexes maintained by independent aggregators. Recent composite intelligence rankings have placed it near the global top ten, just outside the leading tier. Engineers cite strong results on MMLU Pro and AIME 2025, tests covering advanced knowledge, reasoning and mathematics that require several steps of logic.

LG's strategy includes models of different sizes to suit different budgets, alongside development toward multimodal systems that work with text, code and visuals. Multimodal AI processes more than one kind of information, such as combining an image with a written question about it.

Upstage, the startup among the selected teams, is pursuing efficiency through carefully chosen training data and economical training methods. Its Solar Pro 2 has roughly 31 billion parameters and has matched or exceeded larger rivals on several benchmarks. Parameters are the numerical settings a model learns during training.

Solar Pro 2 appears on the Frontier Language Model Intelligence leaderboard and has earned leading positions when intelligence is compared with running cost. That comparison matters to companies paying for repeated AI use. Upstage targets knowledge intensive work in law, finance and healthcare, where accuracy, control and the cost of each task can make a smaller model attractive.

Naver and SK Telecom put Korean language first

Naver trained HyperClova X on one of the largest assembled collections of Korean language material and integrated the model into its search engine and other services. Internal and external tests show an advantage over GPT 4 on tasks involving Korean language, culture and context.

Naver has also released specialized variants and lighter models intended to reduce costs. In June 2025, it introduced HyperClova X Think, designed for more careful reasoning in search and conversation. The company is working toward an omni foundation model that combines text, audio and images, with video also a possible component.

SK Telecom brings its telecommunications customer base and AI investments to the A.X series. Trained from scratch with a strong Korean focus, the models have beaten GPT 4o on local benchmark suites according to company tests. They already power AI secretaries used for customer service, call summaries and other telecommunications tasks.

The company has shared AX 3.1 Lite, a lighter model with roughly 7 billion parameters trained on 1.65 trillion tokens. Tokens are the pieces of text a model processes, which can be words or parts of words. SK Telecom reported about 96 percent performance on KMMLU2 reasoning and 102 percent on the CLIcK3 cultural understanding index relative to larger models.

SK Telecom says its models will be released under open source licenses so developers can build applications across industries. Its consortium plans to release a model this year with a permissive license, allowing broad use under the license's terms.

NC AI joins the field while Kakao stays outside

NC AI, a subsidiary of game publisher NCSoft, surprised many by securing a place in the government program. Its Varco models, including Varco Vision 2.0, show strong image and text processing results even at smaller scales.

The company's focus fits its origins in interactive content rich in graphics. NC AI is assembling a consortium with research groups and smaller companies, drawing on years of internal AI development to compete with larger technology businesses.

Kakao was not selected for the sovereign AI roster, but remains active through its Kanana models and investment in multimodal capabilities. It maintains a mix of smaller and more capable models, with some released under open source terms, for mobile applications and uses involving substantial media content.

Kakao also plans a partnership with OpenAI to bring an AI agent across its platforms. An agent is an AI system designed to carry out tasks, rather than simply return a conversational answer. Kakao's activity means the domestic market extends beyond the five teams receiving government backing.

What the benchmark scores mean for users

AI developers use benchmark suites such as MMLU, HumanEval, LiveCodeBench, AIME and MATH 500 to test knowledge, coding, mathematics and reasoning. These tests help engineers decide how to train models and identify their strengths and weaknesses.

Korean companies are gaining ground on several measures. Exaone and Solar Pro 2 appear on global leaderboards, while SK Telecom and Naver point to advantages on Korean language and culture tests. The national program adds another measure: reaching at least 95 percent of the performance of leading systems released within a rolling period of six months.

Benchmark scores can be close together, and some tests become less useful when many models achieve very high results. Certain tasks reward recalling familiar patterns more than working through new problems. Buyers therefore need to consider how a model performs in the actual work they intend to give it.

OpenAI's GPT 5 has posted gains in mathematics, coding and vision in recent months, while user responses to newer systems can be mixed. Practical usefulness often depends on the application. A model suited to Korean customer calls may offer a different balance of strengths from one used for general coding.

Enterprise buyers also care about control over their data, predictable costs and dependable answers. Korean teams are responding by tailoring models to particular fields and putting them into live services, where performance includes repeated daily use rather than a single test score.

Domestic infrastructure still draws on global hardware

South Korea's semiconductor industry gives the strategy an industrial base. SK Hynix supplies high bandwidth memory used in Nvidia's leading GPUs, while Samsung remains a major memory manufacturer and chip foundry operator.

The model teams still rely on Nvidia hardware for training. SK Telecom plans to use its Titan supercomputer, built with Nvidia GPUs, and an AI data center it is developing with Amazon to train and serve models. The approach combines operations inside Korean facilities with international technology partners.

Domestic neural processing units, or NPUs, offer another route for running trained AI systems. These chips are designed for particular AI calculations. Running a trained model to produce an answer or prediction is called inference.

Rebellions' ATOM chip has entered production use for inference in SK Telecom's X Caliber service, which analyzes animal X rays. It replaced older Nvidia GPU servers in a service used by more than 1,000 veterinary hospitals. X Caliber detects dozens of findings in about 15 seconds and reports accuracy in the high nineties.

FuriosaAI's RNGD chip is used with LG's Exaone 3.5 and received commercialization approval after extensive evaluation. Claims about its performance point to better output per watt than general purpose GPUs for repetitive tasks, making electricity use part of the calculation.

Industry expectations point toward a division of work: GPUs handling flexible training and many other workloads, with NPUs serving targeted inference at large volumes where power consumption and cost matter most. Domestic chips, computing facilities and models could then develop together rather than as separate projects.

Where Korean AI is already entering daily work

Telecommunications companies are using LLMs to transcribe and summarize calls, block spam and assist human customer service agents. Search engines and online portals are improving Korean language results and conversational answers.

Game studios are exploring multimodal systems for character dialogue, visuals and moderation. Financial companies and insurers are testing assistants that retrieve policy details and generate documents that meet compliance requirements. Government agencies see uses for secure domestic systems in public services that require data to stay within the country.

Healthcare is another early focus. In early 2025, South Korea's regulator for food and drugs issued guidance for approving medical AI that generates text, creating a route for clinical applications using LLMs.

Seoul National University Hospital reports that a medical LLM trained on tens of millions of clinical records with identifying information removed exceeded the human average on the Korean Medical Licensing Examination. Its development shows why specialized vocabulary and local medical knowledge matter to Korean applications.

Lighter models such as AX 3.1 Lite are designed to run on devices and at the telecommunications edge, meaning computing facilities closer to users. Processing information nearby can reduce response delays and help protect personal data. Open source access can also lower the barriers for smaller hospitals and startups developing assistants familiar with Korean medical terms.

Open access brings opportunity and debate

Seoul is encouraging open source releases so domestic developers and smaller businesses can adopt, adapt and commercialize the models. Sharing code and model weights where feasible is an expectation accompanying dataset support and GPU leasing. Model weights are the learned numerical settings that allow a trained system to operate.

The government wants this approach to spread AI use across industries and reduce dependence on any single vendor. Naver and LG are developing families that combine proprietary and open releases, balancing commercial products with wider access for developers.

Some startups worry that government backing could favor large conglomerates and squeeze smaller competitors. Others argue that scarce resources would be better directed toward applications for specific industries than toward national foundation models chasing laboratories with enormous budgets.

Policymakers argue that open source terms, shared datasets and staged evaluations will support competition, while a credible national model could create export opportunities. The national AI computing center has nevertheless faced difficulty attracting private partners, showing the need to match infrastructure investment with market demand.

Competition will test more than model quality

OpenAI, Anthropic, Google and xAI are releasing faster and more capable systems while investing heavily in safety and alignment research. Alignment concerns how a model's behavior fits its intended purposes and rules. Chinese companies, including Alibaba and DeepSeek, are also releasing strong open source systems.

Korean models can find advantages at home through language and regulatory fit, but international growth requires attracting developers and demonstrating useful results. Computing budgets and the ability to recruit skilled researchers remain constraints. Government funding, GPU access and international recruitment support address those pressures.

Businesses also need tools to customize models, monitor their behavior and maintain dependable services. Sensitive uses require protection against prompt injection, where instructions attempt to redirect a system, data leakage and hallucinations, which are plausible sounding but incorrect answers.

South Korea's industrial connections offer practical advantages: memory makers supply GPU manufacturers, telecommunications companies operate computing close to users, and chip designers, cloud operators and software teams can learn from one another's deployments. A strong developer community around open models could help turn those connections into a choice for companies and countries seeking alternatives to providers in the United States and China.

Key Points

• Five consortia led by LG AI Research, Naver Cloud, SK Telecom, Upstage and NC AI were selected from 15 applicants.
• The target is at least 95 percent of the performance of leading global models released within the previous six months.
• One team is scheduled to be cut in December 2025, with evaluations every six months and two teams remaining by 2027.
• Support includes datasets worth 10 billion won per team, a broadcast and video data pool valued at 20 billion won, and 2.8 billion won per team for specialized datasets.
• LG and Upstage compete on reasoning and efficiency; Naver and SK Telecom focus strongly on Korean language; NC AI brings image and text capabilities.
• Kakao remains active outside the program through Kanana and a planned OpenAI partnership.
• The strategy combines domestic models, cloud services and chips with global hardware partnerships.
• Open source access is intended to support wider adoption, while funding, reliability and competition remain execution challenges.