South Korea Backs Five AI Teams in Push to Rival ChatGPT on Home Ground

Five teams, a 95 percent target and a narrowing race

South Korea has selected five industry consortia to build an independent artificial intelligence foundation model, setting a target of at least 95 percent of the performance of leading systems such as ChatGPT. Led by LG AI Research, Naver Cloud, SK Telecom, Upstage and NC AI, the teams are competing to deliver technology that runs on local infrastructure and understands Korean language and culture.

As of September 2, 2025, the Ministry of Science and ICT (MSIT) has moved the Independent AI Foundation Model project into execution. The five teams were chosen from 15 applicants, meaning one in three applicants advanced. Their consortia combine companies, universities, research institutes and government bodies.

Each team will receive processed datasets valued at 10 billion won, broadcast video and training data valued at 20 billion won, and 2.8 billion won in direct funding for custom datasets. Those three categories together represent 32.8 billion won in data support per team. SK Telecom and Naver Cloud also plan to supply leased graphics processing unit infrastructure to support training for several teams.

The government plans to review progress every six months, reducing the field from five teams to four in December and eventually to two by 2027. The central challenge is practical as well as technical: Korea wants models that can compete internationally while serving local businesses and public services accurately, affordably and under domestic control.

A foundation model is a broadly trained AI system that can become the base for many applications, from answering questions to writing software. Korea's project seeks to build that base at home, rather than depending entirely on systems developed abroad.

What sovereign AI means for Korea

The administration describes sovereign AI as autonomy across the full technology chain, covering chips, cloud computing, models and services. Its strategy responds to the concentration of AI leadership in the United States and China, alongside concerns about data security, language accuracy and cultural understanding.

Korea brings established strengths in semiconductors, telecommunications networks and applied AI. Policymakers argue that these capabilities can support a distinct national approach, while adoption of open source technology can allow businesses across the economy to build on domestic tools.

The project also has a staged schedule, with infrastructure provision and competitive reviews running alongside model development:

• June 2025: Naver introduced HyperClova X Think, expanding its reasoning capabilities.
• September 2, 2025: Five selected consortia are participating in the national foundation model effort.
• Late 2025 through mid 2026: SK Telecom and Naver Cloud plan to provide leased GPU infrastructure for training.
• December 2025: The government plans to reduce the five participating teams to four.
• By 2027: Two teams are expected to remain in the program.

The 95 percent performance target provides a headline goal, but the project description does not specify a single benchmark or model version against which that threshold will be measured. The teams are pursuing several kinds of capability, including reasoning, image understanding and Korean language performance.

Why Korean language processing can cost more

Naver's HyperClova team points to a financial reason for domestic models: some systems designed primarily around English require more processing to handle Korean. An analysis cited by the team estimates that processing Korean can carry a cost burden 347 percent higher than processing English on some such models.

Language models break text into small units called tokens, then learn patterns across those units. An English word may fit into one token. Korean and other languages written outside the Latin alphabet can be divided into more pieces, including units based on bytes or parts of words. More tokens mean longer sequences to process, greater computing demand and higher customer bills.

A tokenizer designed for Korean, combined with training material that reflects actual Korean usage, can improve speed and accuracy while reducing that burden. Naver presents its domestic model and tools as an economic response to the costs created by English focused tokenization and pricing.

Korean cultural knowledge is another concern. Popular chatbots have produced confident errors about Korean history and public figures, and have struggled with familiar jokes and memes. Those failures weaken trust when users expect an assistant to understand local references. Training on Korean material and aligning models with Korean usage can reduce such mistakes.

LG and Naver pursue reasoning and visual understanding

LG AI Research's Exaone 4.0 combines reasoning with visual understanding. LG offers versions aimed at high performance and lighter deployment requirements, reflecting a wider shift among Korean developers from simply increasing model size toward usefulness in actual services.

Exaone 4.0 has recorded competitive results on global indexes and appeared near leading systems on widely watched leaderboards. An expert variant with roughly 32 billion parameters was credited with passing several Korean national qualification examinations. Parameters are the internal numerical settings a model learns during training; their number describes model scale, while performance also depends on data and training design.

LG's business focus includes document understanding, coding assistance and visual analysis for industrial settings. Its cooperation with domestic chip designers also connects model development to hardware: FuriosaAI's RNGD accelerator was adopted to run an earlier Exaone release with strong performance per watt.

Naver Cloud's HyperClova family draws on one of the largest Korean language collections assembled for AI training. HyperClova X is tuned for accuracy on Korean tasks and includes specialized variants. HyperClova X Think, introduced in June 2025, strengthens reasoning and can solve mathematics and science problems containing images.

Naver also plans a multimodal omnifoundation model spanning text, images and other media. Multimodal systems work with more than one kind of information, allowing a model to interpret a picture alongside a written question. For Naver, the technical ambition sits alongside the goal of handling Korean history, law and cultural context more faithfully.

SK Telecom and Upstage put efficiency at the center

SK Telecom's AX models focus on Korean language performance, efficiency and integration into telecommunications products. AX 3.1 Lite has 7 billion parameters and was trained from scratch on about 1.65 trillion multilingual tokens, with a heavy Korean focus. It is available as open source for mobile applications and use directly on devices.

Company benchmarks report relative performance of about 96 percent on the Korean reasoning test KMMLU2 and 102 percent on the cultural understanding benchmark CLIcK3 compared with larger models. These figures describe performance relative to comparison models, rather than the percentage of questions answered correctly.

SK Telecom has tested call summaries, spam filtering and assistant features, and is deploying AI secretaries for consumers. Its participation in the national project extends beyond models and applications to providing GPU capacity for other teams.

Upstage is pursuing what it calls intelligence per dollar. Its Solar Pro 2, with roughly 31 billion parameters, is the only Korean entry on Artificial Analysis's Frontier Language Model Intelligence leaderboard. It ranks near the global top tier for intelligence and leads the Intelligence versus Cost to Run metric, which compares capability with the computing cost needed to deliver it.

Upstage targets enterprise uses in law, finance and healthcare while maintaining multilingual capabilities. Its argument is that a carefully trained model with around 30 billion parameters can match or surpass much larger competitors on many tasks when developers choose an effective balance between training tokens and model size.

Gaming, messaging and specialist models widen the field

NC AI, a subsidiary of NCSoft and one of the five selected consortium leaders, is developing the multimodal Varco family. Varco and Varco Vision 2.0 process images and text and have outperformed other open source models on several tests. Their focus includes media, games and creative tools, where interpreting images is central to the work.

Kakao was not selected for the government project but continues to develop its Kanana models in partnership with OpenAI. It is integrating assistants into messaging and content platforms used daily by tens of millions of Koreans. That distribution gives Kakao a substantial role in bringing AI into everyday consumer services.

KT has introduced MIDEUM 2.0, an open source model with 11.5 billion parameters, and plans a Pro version with stronger reasoning. These efforts show that the national competition sits within a broader industry push involving both major companies and smaller specialists.

Seoul National University Hospital trained a Korean clinical model on tens of millions of records. It reported a score of 86.2 percent on the Korean Medical Licensing Examination, above the human average in a comparative study. Hospital researchers are using locally relevant medical data to pursue near expert performance on licensing tests.

Mathpresso and Upstage jointly developed MATH GPT, a model with 13 billion parameters. It scored 0.488 on one mathematics benchmark against GPT 4's 0.425 while using much less computing power. The difference was 0.063 on that benchmark, illustrating why developers are investing in focused training rather than relying only on larger models.

What the benchmark scores tell buyers

AI benchmarks test general knowledge, coding, mathematics and reasoning. Common suites include MMLU, HumanEval, LiveCodeBench, AIME and MATH 500. More demanding tests, including MMLU Pro and Humanity's Last Exam, raise the difficulty as model capabilities improve.

Hugging Face and Artificial Analysis aggregate results into leaderboards that help shape developers' reputations and influence purchasing decisions. Exaone 4.0's proximity to leading systems and Solar Pro 2's frontier placement give Korean developers visible measures of progress.

Solar Pro 2's position on Intelligence versus Cost to Run is especially relevant when GPU supply is tight and businesses must control cloud inference spending. Inference is the stage when a trained model processes a user's request and generates an answer. Its cost matters repeatedly as services gain customers.

Businesses also need reliability, an appropriate tone and a good fit for the task. Leading frontier models have faced user criticism for bland conversation and basic mistakes despite strong academic test results. Korean models will face the same everyday judgment: whether people choose to use their services in Korean.

Domestic chips move into working services

SK Telecom has begun replacing Nvidia GPU servers with Rebellions' ATOM neural processing units for X Caliber, its veterinary X ray analysis service. The service is used by more than 1,000 hospitals nationwide and detects 45 categories of findings, including markers related to breed.

X Caliber returns results in roughly 15 seconds, with reported accuracy reaching 97 percent. SK Telecom built a production cluster of six servers, each containing 48 NPUs, for a total of 288 units. The company is also piloting NPUs for assistant features in consumer services.

NPUs are chips designed for neural network workloads. They can reduce electricity use and cost for repeated inference tasks in healthcare, education and telecommunications. X Caliber provides a practical example of domestic hardware operating in a service already used by hospitals.

FuriosaAI's second generation RNGD chip has been integrated with LG's Exaone 3.5 for inference and received approval for commercialization after extensive evaluation. FuriosaAI has also announced its first overseas customer in the United States, extending its business beyond domestic deployments.

GPUs remain essential for general purpose training and many inference workloads, so Korea's infrastructure is likely to combine GPUs and NPUs. Telecommunications operators are pairing data center deployments with edge computing, which places processing closer to users to reduce response delays, in line with national efforts to upgrade AI infrastructure across the country.

Where people will encounter Korean AI

Healthcare combines several parts of the strategy: locally trained medical models, domestic chips and rules for clinical use. Early in 2025, the regulator issued guidance for approving systems that generate text in clinical settings, giving hospitals and vendors a route to validate new tools safely.

Telecommunications companies are bringing language models into consumer applications and contact centers. SK Telecom's assistants can summarize calls, flag unwanted traffic and help recommend plans. Naver is improving search, shopping and customer support with HyperClova variants designed to understand Korean wording and cultural context.

Kakao is introducing AI agents across messaging and content services, where convenient integration can matter more to users than model size. Several laboratories are also releasing smaller models for use directly on devices. Handling simple tasks locally can reduce delays and protect privacy.

Legal and financial firms want systems that cite statutes correctly and draft documents using specialist knowledge. Startups and universities are targeting these needs, while compact mathematics models show how focused reasoning systems can compete with much larger rivals using a fraction of the computing resources.

The practical direction is consistent across these applications: developers are selecting data and designing models for particular jobs. A useful assistant must fit the work of a hospital, customer service team or legal office, rather than simply produce a strong general benchmark score.

Funding choices and the debate over national champions

The government has discussed an investment plan of 100 trillion won over several years and appointed industry figures to senior positions. They include an AI senior presidential secretary with a background in Naver's research organization.

Two major programs anchor the strategy. The Independent AI Foundation Model project has attracted strong private participation. The National AI Computing Center, intended to pool and allocate computing capacity for researchers and industry, has progressed more slowly, reflecting the difficulty of assembling infrastructure consortia.

Some startup founders worry that funding and attention will flow mainly to large conglomerates. Critics favor concentrating on specialized applications where Korean companies can succeed quickly, rather than competing directly with global technology giants on broad foundation models.

Supporters argue that domestic models are needed for data protection, compliance with sector rules and cultural fit. Open source licensing, they contend, can spread the benefits beyond the companies leading development and allow other businesses to build services on shared technology.

Korea's leaders also present sovereign AI as compatible with international cooperation. Opinion leaders have called for national control alongside cooperation on global safety and transparency norms. The country's semiconductor strengths could help it serve partners seeking local control without withdrawing from global trade. Korea's experience will offer other middle powers a concrete example of how national autonomy and international cooperation can work together.

Key Points

• MSIT selected five consortia led by LG AI Research, Naver Cloud, SK Telecom, Upstage and NC AI from 15 applicants.
• The project targets at least 95 percent of leading systems' performance, with Korean language capability and local infrastructure central to its goals.
• Each team is allocated data support totaling 32.8 billion won across three categories.
• The field is scheduled to narrow to four teams in December 2025 and two by 2027.
• LG, Naver, SK Telecom and Upstage are pursuing reasoning, visual understanding, Korean accuracy and computing efficiency.
• NC AI, Kakao, KT, hospitals and specialist developers extend the effort into games, messaging, medicine and mathematics.
• Rebellions and FuriosaAI chips are entering commercial services, while GPUs remain essential for training and many applications.
• Korea's policy debate centers on how to balance national model development, specialist applications, broad access and international cooperation.