From major corporations to emerging startups, South Korean companies are crafting large language models that reflect their unique language and cultural context, positioning themselves to rival international leaders such as OpenAI and Google.
In the previous month, South Korea unveiled its boldest national AI project so far, committing ₩530 billion (approximately $390 million) to support five domestic firms working on large foundational models.
This initiative highlights Seoul’s intention to reduce dependence on overseas AI solutions, aiming to bolster national security and maintain stronger oversight of data in the age of artificial intelligence.
The Ministry of Science and ICT selected LG AI Research, SK Telecom, Naver Cloud, NC AI, and the startup Upstage to participate in the competition.
Every half year, the government will assess the progress of these initial participants, eliminate those falling behind, and continue to fund the top performers until only two remain to spearhead the country’s sovereign AI ambitions.
Each contender brings a unique strength to South Korea’s AI development. TechCrunch interviewed several of the chosen companies to learn how they intend to challenge OpenAI, Google, Anthropic, and others on their own ground. NC AI opted not to provide comments.
LG AI Research: Exaone
LG AI Research, the research and development division of LG Group, has introduced Exaone 4.0, a hybrid reasoning AI model. This latest iteration merges comprehensive language understanding with advanced reasoning capabilities that debuted in the earlier Exaone Deep model.
Exaone 4.0 (32B) already performs competitively with other models on Artificial Analysis’s Intelligence Index benchmark, as does Upstage’s Solar Pro2. LG aims to climb even higher by leveraging extensive access to real-world industry data, spanning sectors from biotechnology to advanced materials and manufacturing.
They are combining this data with a focus on refining it before using it to train their models. Rather than simply scaling up, LG is prioritizing smarter processes so its AI can deliver tangible, practical benefits beyond what general-purpose models provide. “This is our core philosophy,” co-head Honglak Lee shared with TechCrunch.
LG is enhancing its models through established methods: making them available via APIs and then using the real-world data generated by users to further train and improve the models.
“As LG’s models advance, our partners are able to provide improved services, which in turn generate more economic value and richer datasets,” he explained.
Yet, instead of pursuing vast GPU clusters, LG AI Research is concentrating on maximizing efficiency, optimizing each chip’s performance, and developing models tailored to specific industries, he added. The aim is not to outspend global competitors, but to surpass them with high-performing, efficient AI solutions.
SK Telecom: A.X
SK Telecom (SKT), South Korea’s leading telecom provider, launched its personal AI assistant A. (pronounced A-dot) in late 2023 and introduced its new large language model, A.X, this July.
A.X 4.0, based on Alibaba Cloud’s open-source Qwen 2.5, is available in two versions: a robust 72-billion-parameter model and a more compact 7B variant.
According to SK, A.X 4.0 handles Korean language inputs about 33% more efficiently than GPT-4o, highlighting its advantage in local language processing. (There is no available comparison with OpenAI’s GPT 5.0.) Earlier this year, SKT also released its A.X 3.1 models as open source. The A. service itself features AI-powered call summaries and automated note-taking, and by August 2025, it had attracted roughly 10 million users.
SK’s strength lies in its adaptability, thanks to access to a wide range of information from its telecom network, including navigation and taxi services.
“SK Telecom’s mission is to connect advanced model research with real-world applications. Leveraging our telecom infrastructure, large user base, and proven services like A., we integrate AI into daily life, from customer support to transportation and manufacturing,” said Taeyoon Kim, head of the foundation model office at SK Telecom, in an interview with TechCrunch.
SK Telecom is also expanding its AI infrastructure, utilizing GPUaaS—the country’s largest GPU-based service—and constructing a new hyperscale AI data center in partnership with AWS. Where resources are lacking, the company seeks strategic partnerships.
“We’re developing a comprehensive ecosystem in collaboration with Korean AI chipmaker Rebellions, securing reliable data partnerships with government and academic institutions, and building a global research network,” Kim stated. “This includes initiatives like our partnership with MIT (MGAIC), which applies foundational models to advanced manufacturing, battery, and semiconductor innovation.”
Naver Cloud: HyperCLOVA X
Naver Cloud, the cloud division of South Korea’s top internet company, launched its large language model HyperClova in 2021. Two years later, it introduced HyperCLOVA X, along with new AI-powered offerings: CLOVA X, a chatbot, and Cue, a generative AI search engine positioned as a competitor to Microsoft’s CoPilot-enhanced Bing and Google’s AI Overview. This year, it also debuted its multimodal reasoning model, HyperCLOVE X Think.
According to a Naver spokesperson, the real value of LLMs lies in their ability to act as “connectors,” bridging legacy systems and isolated services to enhance overall utility.
Naver distinguishes itself as the only Korean company—and among the few globally—to possess a true “AI full stack.” It has developed its HyperCLOVA X model from the ground up and operates the extensive data centers, cloud infrastructure, AI platforms, applications, and consumer services that bring the technology to users, the spokesperson noted.
Much like Google, but tailored for South Korea, Naver is integrating its AI into essential services such as search, e-commerce, maps, and finance. Its key advantage is access to real-world data. For example, its AI Shopping Guide provides recommendations based on actual consumer preferences. Other offerings include CLOVA Studio, which enables businesses to create customized generative AI, and CLOVA Carecall, an AI-driven check-in service designed for seniors living alone.
The Naver spokesperson emphasized that surpassing global AI leaders like OpenAI and Google depends on perfecting its model “recipe” and securing sufficient investment to scale. Nevertheless, the company prioritizes sophistication over sheer size, asserting that its AI is already globally competitive at similar scales.
Upstage’s Solar Pro 2
Upstage is the sole startup involved in this initiative. Its Solar Pro 2 model, released last July, became the first Korean model to be recognized as a frontier model by Artificial Analysis, placing it alongside OpenAI, Google, Meta, and Anthropic, according to Soon-il Kwon, executive vice president at Upstage.
While most frontier models have between 100 billion and 200 billion parameters, Solar Pro 2 contains just 31 billion, yet delivers superior performance for Korean users and is more cost-efficient, Kwon told TechCrunch.
“Solar Pro 2 has outperformed international models on key Korean benchmarks. Through this project, Upstage aims to reach 105% of the global standard in Korean language performance,” Kwon stated.
Upstage seeks to set itself apart by prioritizing real business outcomes rather than just benchmark scores, he said. The company is developing specialized models for sectors such as finance, law, and healthcare, and is working to foster a Korean AI ecosystem driven by “AI-native” startups.
