From high-tech giants to startups, Korean players are developing large-scale language models tailored to their own language and cultures that are ready to compete with global heavyweights such as Openai and Google.
Last month, the country launched its most ambitious sovereign AI initiative to date, pledging 500 billion (approximately $390 million) to five local businesses building a massive foundation model.
The move highlights Seoul’s desire to reduce dependence on foreign AI technologies, hoping to strengthen national security and maintain tighter control of data in the AI era.
The organizations chosen by the Ministry of Science to compete were LG AI Research, SK Telecom, Naver Cloud, NC AI, and Startup Upstage.
Every six months, the government will check the progress of the first cohort, cut underperformers, and continue to fund the frontline until only two people are left to lead the country’s sovereign AI drive.
Each player brings different benefits to the Korean AI race. TechCrunch spoke about some of the selected companies and how Openai, Google, humanity and others deal with it on their grass at home. NC AI declined to comment.
LG AI Research: Exaone
LG AI Research, the R&D unit of the Korean giant LG Group, offers the hybrid inference AI model, Exaone 4.0. The latest version blends a wide range of language processing with advanced inference capabilities first introduced in the company’s previous ExaONE deep model.
TechCrunch Events
San Francisco
|
October 27th-29th, 2025
EXAONE 4.0 (32b) has already scored appropriately against competitors in the AI-analytic Intelligence Index Benchmark (similar to Upstage’s Solar Pro2). But it plans to improve and increase the rank through deep access to real-world industry data, from biotechnology to advanced materials and manufacturing.
It combines data focused on improving the data before feeding it to the model to train. Instead of chasing pure scale, LG wants to make the entire process more intelligent, so its AI can provide real and practical value beyond what a generic model can offer. “This is our basic approach,” co-head Honrak Lee told TechCrunch.
LG is improving its model through familiar tactics. Provide them through APIs and train and improve the model using actual data generated by users of those services.
“As our LG model improves, our partners can provide better services, which can generate greater economic value and richer data,” he said.
However, instead of chasing large GPU clusters, LG AI research focuses on efficiency, making the most of all chips and creating industry-specific models, he said. The goal is not to outperform the global giants, but to cover them with high-performance, efficient AI.
Korean telephone company Giants SK Telecom (SKT) launched the Personal AI Agent A (pronounced A-Dot) service in late 2023, and launched AX, a new major language model in July this year.
Built on top of Alibaba Cloud’s Chinese open source model, Qwen 2.5, AX 4.0 comes in two models: 7.2 billion parameter versions and a lightweight 7B version.
SK says that AX 4.0 handles Korean inputs about 33% more efficiently than GPT-4O, highlighting the edge of local languages. (Openai’s GPT 5.0 comparison data is not available.) SKT sourced the AX 3.1 model early this summer. Meanwhile, A. The service provides features such as AI call summary and auto-generated notes. As of August 2025, it has already attracted around 10 million subscribers.
The edge of SK is because information can be accessed from telecom networks ranging from navigation to taxi hailing.
“The role of SK Telecom is to act as a bridge between cutting-edge model research and real-world impact. With telecom infrastructure, extensive user base and proven services like AA, we bring AI directly into both customer service, mobility and manufacturing.”
SK Telecom is investing in AI infrastructure using GPUAAS, Korea’s largest GPU-based service, to build a new hyperscale AI data center with AWS. Whatever it is missing, it is partnered to get it.
“We are building a full-stack ecosystem with the uprising of South Korean AI chip makers, ensuring reliable data partnerships through work with governments and universities, and fostering a global research network,” Kim said. “This includes projects such as collaboration with MIT (MGAIC), which applies the fundamental model to advanced manufacturing and battery and semiconductor innovation.”
Naver Cloud: Hyperclova x
Naver Cloud, the cloud services division of a leading Korean internet company, introduced a large-scale language model, Hyperclova, in 2021. Two years later, we announced the upgraded version, the Hyperclova X. An overview of AI. This year we also announced the Hyperclove X Think, a multimodal inference AI model.
Naver Cloud believes that the true power of LLMS is to act as a “connector” that links legacy systems with siloed services to improve their usefulness.
Naver stands out as the only company in Korea and one of the few companies in the world, and can really claim to have a “AI full stack.” It builds Hyperclova X models from scratch, implements large data centers, cloud services, AI platforms, applications and consumer services, and delivers technology, the spokesman explained.
Similar to Google, but tweaked for Korea, Naver embeds AI into core services such as search, shopping, maps and finance. The advantage is the actual data. For example, AI Shopping Guides provide recommendations based on what people actually want to buy. Other services include Clova Studio. ClovaStudio allows businesses to build custom-generated AI.
A Naver spokesperson says it has acquired a global AI Giants like Openai and Google has two things. The “recipe” is completed for the model and is fixed to secure the capital. Still, rather than chasing size, the company emphasizes sophistication, claiming that AI is already globally competitive on comparable scales.
Solar Pro 2 from Upstage
Upstage is the only startup that competes in a project. The Solar Pro 2 model, released last July, was recognized as a frontier model by artificial analysis, and the first Korean model to be put into the ring alongside Openai, Google, Meta and mankind, according to Soon-Il Kwon, Executive Vice President of Upstage.
While most frontier models have 100-200 billion parameters, the Solar Pro 2 has only 31 billion parameters, making it better performance for Koreans and is more cost-effective, Kwon told TechCrunch.
“The Solar Pro 2 is superior to the global model on the major Korean benchmarks. With this project, Upstage aims to achieve Korean performance at 105% of the global standard,” Kwon said.
Upstage aims to distinguish itself by focusing on its actual business impact, not just benchmarks, he said. Therefore, we are promoting the creation of the Korean AI ecosystem led by AI-Native startups, while developing specialized models for industries such as finance, law, and medicine.
