South Korean players, from tech giants to startup companies, are developing large language model tailored to their language and culture to compete with global heavyweights such as OpenAI and Google.
Last monththe nation launched its largest sovereign AI initiative to-date, pledging to W530 billion (about $390 millions) to five local firms building large-scale, foundational models.
Seoul is attempting to reduce its reliance on foreign AI technology, in order to improve national security and maintain a tighter grip on data.
These are the organizations selected by the Ministry of Science and ICT (19459058) to compete LG AI Research SK Telecom () Naver Cloud, Nc aiand the startup Upstage
The government will review every six months the progress of the first cohort, cut the underperformers and continue funding the leaders until only two remain to lead South Korea’s autonomous AI drive.
Every player brings a unique advantage to South Korea’s AI race. TechCrunch interviewed several of the selected firms about how they plan on taking on OpenAI and Google on their own turf. NC AI declined comment.
Exaone
LG AI Research is the R&D division of South Korean giant LG Group. Exaone 4.0 is a hybrid AI model that combines reasoning and broad language processing. The latest version combines broad language processing and advanced reasoning features introduced in the company’s earlier Exaone Deep models.October 27-29, 2025
Exaone 4.0 (32B) already scores reasonably well against competitors on Artificial Analysis‘s Intelligence Index (as well as Upstage’s) benchmark Solar Pro2). It plans to improve its ranking and to move up through its access to real-world data from industries ranging from advanced materials to biotech.
It is combining that data with a focus to refine the data before feeding it to the models to be trained. LG is not interested in chasing scale but instead wants to make its AI more intelligent. This will allow it to deliver real value beyond what general-purpose AI models can provide. Honglak Lee, co-head of LG’s AI division, told TechCrunch that this is the company’s fundamental approach.
LG improves its models using familiar tactics: it offers them via APIs and then uses the real-world data produced by users to train the model. “As LG improves its models, our partners will be able to deliver better services which, in turn, will generate greater economic value, and even richer information,” he said.
Instead of chasing massive GPU Clusters, LG AI Research focuses on efficiency, getting more out of each chip, and creating industry specific models, he said. The goal is not to outspend global giants, but to outsmart with high-performance, yet more efficient AI.
South Korea’s telco giant SK Telecom launched its Personal AI agent A.service (pronounced A-dot), way back in 2023, and just rolled its new large language A.X model this July.
Built atop the Chinese open-source model from Alibaba Cloud. Qwen 2.5 (19459058),””https://huggingface.co/skt/A.X-4.0?utm_source=chatgpt.com” ” rel=””noreferrer noopener nofollow”” target=””_blank” “> A.X 4.0 is available in two versions, a heavy 72-billion parameter version and a lighter version 7B.
SK claims that A.X 4.0 processes Korean entries about 33% faster than GPT-4o, highlighting its local language advantage. OpenAI’s GPT5.0 comparison data is unavailable. SKT has also opened sourced its A.X 3.1 models (19459058) were released earlier this summer. The A. service, meanwhile, offers features such as AI call summaries or auto-generated notes. By August 2025, the service will have attracted about 10 million subscribers.
SK has a competitive edge because of its versatility. It can access information from its telecom network ranging anywhere from navigation to taxi-hailing.
SK Telecom is a bridge between cutting edge model research and real world impact. Taeyoon KIM, the head of SK Telecom’s foundation model office, told TechCrunch that with our telecom infrastructure, large user base, and proven services like A. we bring AI into everyday life.
SK Telecom also invests in AI infrastructure. It uses GPUaaS, South Korea’s largest GPU-based services, and builds a new hyperscale AI Data Center with AWS. It is partnering with others to acquire what it lacks.
Kim said, “We are building a full stack ecosystem with Korean AI chips maker Rebellions. We are securing trusted partnerships through work with government and universities and fostering global research networks.” “That includes projects such as our collaboration with MIT’s (MGAIC), that applies foundation models to advanced production, battery and semiconductor innovation.
Naver Cloud: HyperCLOVA
In 2021, Naver Cloud, a cloud services division of South Korea’s largest internet company, launched its large language model HyperClova. HyperCLOVA X was released two years later along with new products powered using the technology. CLOVA X is an AI chatbot and Cue is a generative AI search engine positioned to compete with Microsoft’s CoPilot enhanced Bing and Google AI Overview. It also revealed this year its multimodal AI model. HyperCLOVE ThinkNaver Cloud believes that LLMs are most powerful when they act as “connectors”linking legacy systems to siloed service to improve their usefulness.
Naver is the only company in Korea — and one of few in the entire world — to be able to claim to have a “AI full stack”. It has built its HyperCLOVA model from scratch, and manages the massive data centers and cloud services, AI platforms and applications, as well as consumer services which bring the technology to live, explained the spokesperson.
Naver, which is similar to Google, but tailored for South Korea, is embedding AI into core services such as search, shopping and finance. Its advantage is the real-world data. Its AI Shopping Guide offers recommendations based upon what people want to buy. CLOVA Studio allows businesses to build custom generative AI. CLOVA Carecall is an AI-powered check in service for seniors who live alone.
According to the Naver spokesperson, beating global AI giants such as OpenAI and Google depends on two things: perfecting their “recipe” of models and securing capital to scale them. The company, however, is not chasing scale, but rather focuses on sophistication, arguing that its AI is already global competitive at comparable sizes.
Upstage’s Solar Pro 2 (19659028) Upstage is the sole startup competing in this project. According to Soon-il Kwon of Upstage, its Solar Pro 2 model was launched in July and was the first Korean model to be recognized as a Frontier Model by Artificial Analysis . It is now competing with OpenAI, Google Meta, Anthropic and OpenAI. Solar Pro 2 has only 31 billion parameters, compared to the 100 billion-200 billion parameters of most frontier models. This makes it more cost-effective and better for South Koreans, according to Kwon, executive vice president at Upstage.
Solar Pro 2 has outperformed the global models in major Korean benchmarks. Upstage’s goal is to achieve a Korean performance that is 105% the global standard with this project,” Kwon said.
Upstage wants to differentiate itself from other companies by focusing not only on benchmarks but also on the real business impact. It is creating specialized models for sectors like finance, law and medicine while also pushing to build a Korean AI eco-system led by “AI native” startups.
