South Korean companies, from major tech firms to startups, are actively developing large language models specifically designed for their own language and culture. These models are poised to compete with global leaders like OpenAI and Google. Last month, the country launched its most ambitious sovereign AI initiative to date, pledging 530 billion won, approximately 390 million US dollars, to five local companies tasked with building large-scale foundational models. This move highlights Seoul’s goal to reduce reliance on foreign AI technologies, aiming to strengthen national security and maintain tighter control over data in the AI era.
The Ministry of Science and ICT selected five organizations to compete: LG AI Research, SK Telecom, Naver Cloud, NC AI, and the startup Upstage. The government will review the progress of this first group every six months, cutting underperformers and continuing to fund the frontrunners until only two companies remain to lead the nation’s sovereign AI initiative.
Each selected company brings a distinct advantage to South Korea’s AI race.
LG AI Research, the research and development unit of the LG Group, offers the Exaone 4.0 model, a hybrid reasoning AI. The latest version combines broad language processing with advanced reasoning capabilities. While Exaone 4.0 already scores reasonably well on industry benchmarks, the company plans to improve its ranking by leveraging its deep access to real-world industry data from fields like biotech, advanced materials, and manufacturing. Instead of focusing on sheer scale, LG aims to make the AI development process more intelligent by refining data before training, with the goal of delivering practical value beyond general-purpose models. The company is improving its models by offering them through APIs and using the real-world data generated by users for further training. LG AI Research emphasizes efficiency, focusing on getting the most from each computer chip and creating industry-specific models, intending to outsmart global giants with high-performing, efficient AI rather than trying to outspend them.
SK Telecom, the telecommunications giant, launched its personal AI agent called A. in late 2023 and introduced its new large language model, A.X, in July. Built on an open-source model from Alibaba Cloud, A.X 4.0 comes in two versions, one with 72 billion parameters and a lighter version with 7 billion. SKT states that A.X 4.0 processes Korean inputs about 33 percent more efficiently than GPT-4o, underscoring its local language advantage. The A. service, which includes features like AI call summaries, has already attracted about 10 million subscribers. SKT’s strength lies in its versatility, with access to data from its telecom network, including navigation and taxi-hailing services. The company sees its role as a bridge between model research and real-world impact. SK Telecom is also investing in AI infrastructure, using South Korea’s largest GPU-based service and building a new hyperscale AI data center with AWS. The company is building a full-stack ecosystem through partnerships with AI chipmaker Rebellions, securing data partnerships with the government and universities, and fostering a global research network, including a collaboration with MIT.
Naver Cloud, the cloud services arm of South Korea’s leading internet company, introduced its HyperClova model in 2021 and unveiled an upgraded version, HyperCLOVA X, two years later. This technology powers products like the CLOVA X chatbot and the Cue search engine. The company believes the true power of large language models is to act as connectors that link legacy systems and siloed services. Naver stands out as one of the few companies in the world with a complete AI stack, having built its model from scratch and operating the data centers, cloud services, platforms, and applications that bring the technology to life. Similar to Google but tailored for South Korea, Naver is embedding its AI into core services like search, shopping, maps, and finance. Its advantage comes from real-world data, which powers services like its AI Shopping Guide. The company states that competing with global giants depends on perfecting its model recipe and securing capital for scaling, but emphasizes sophistication over sheer size.
Upstage is the only startup in the government project. Its Solar Pro 2 model, launched last July, was the first Korean model recognized as a frontier model by Artificial Analysis, placing it alongside models from OpenAI, Google, Meta, and Anthropic. While most frontier models have 100 to 200 billion parameters, Solar Pro 2, with only 31 billion parameters, performs better for South Koreans and is more cost-effective. The model has outperformed global models on major Korean benchmarks, and with the government project, Upstage aims to achieve a Korean language performance that is 105 percent of the global standard. The company differentiates itself by focusing on real business impact, developing specialized models for industries such as finance, law, and medicine, and working to build a Korean AI ecosystem led by AI-native startups. NC AI declined to comment on its plans.