September 3, 2025
SEOUL – South Korea is muscling into the global competition for large language models, with both conglomerates and startups rolling out artificial intelligence systems tailored to Korean language and culture, as they take aim at OpenAI and Google’s dominance.
From LG’s Exaone and Naver’s HyperClova X to Upstage’s Solar Pro, SK Telecom’s A.X series, Kakao’s Kanana and NC AI’s Varco, Korean developers are jockeying for a spot in the government’s “sovereign AI” drive.
Five consortia led by LG AI Research, SKT, Naver, NC AI and Upstage have been tapped to participate in the initiative that seeks to deliver a sovereign AI foundation model. Though their models have tens of billions of parameters, compared to the trillions that the leading LLMs have, the project aims to reach at least 95 percent of the performance of frontier models such as ChatGPT. Ultimately, two final teams will be chosen through a series of evaluations.
The Korea Herald looks into how these players stack up and how other challengers are trying to carve out space in Korea’s fast-evolving LLM landscape.
LG AI Research’s Exaone 4.0, unveiled in July, has quickly become Korea’s flagship contender in the sovereign AI race. It was one of only two Korean LLMs to be listed on the Artificial Analysis Intelligence Index — a prominent index evaluating global AI models. It ranked first among Korean models, fourth among open-weight models and 11th overall worldwide in the Intelligence Index at the time of release.
The hybrid LLM, which incorporates a reasoning function, performs competitively against some frontier models that are several times larger in scale.
The system is offered in two formats: a high-performance 32B model — indicating 32 billion parameters — for computation-heavy tasks and a lightweight 1.2B version optimized for on-device use. It also has visual analysis capabilities that allow it to work with unstructured data such as images, documents and charts. On the specialized “Chart QA” benchmark, which measures precision in interpreting corporate documents and charts, Exaone achieved one of the highest scores globally.
LG’s LLM is largely focused on enterprise needs such as automating workflows, analyzing data and supporting research in industries like finance, pharmaceuticals and manufacturing.
Selected among the five in the government national AI competition, LG aims to develop the world’s highest-level frontier AI model, dubbed “K-Exaone,” based on its experience of developing Exaone 4.0 from scratch.
Telecommunication giant SK Telecom is making its own sovereign AI bid with A.X, an open-source model tailored specifically to Korean language and culture.
Developed on top of the Chinese Qwen 2.5 base and expanded in-house, A.X 4.0 has posted stronger results than OpenAI’s GPT-4o on local benchmarks, scoring 78 on KMMLU and 83 on CLIcK — tests that measure knowledge and cultural nuance in Korean.
Its A.X 3.1 model, released in July, was developed with a “from scratch” approach using its new proprietary technology with 34 billion parameters. SKT said it developed it from scratch, covering all stages of the process, including model design and data training.
Since releasing KoBERT in 2019, SKT has steadily advanced its own models through both large-scale continued pretraining. The latest offerings, A.X 4.0 and A.X 3.1, have shown performance comparable to OpenAI’s GPT-4o, while demonstrating world-class ability in understanding Korean-language context, according to the company.
The company is already putting the model into practice, offering it as an “AI secretary,” that summarizes customer-service calls. With reasoning and multimodal upgrades scheduled later this year, the mobile carrier plans to turn it into a paid service.
Beyond consumer services, SKT is pitching A.X as a sovereign option for government agencies and financial institutions that prefer on-premises systems over foreign cloud platforms.
The model’s inclusion in the national AI competition underscores how sovereign AI in Korea is not confined to labs; it is about integrating homegrown systems into telecom infrastructure and everyday services.
South Korean AI startup Upstage is challenging global giants with Solar Pro 2, a lean but powerful model launched in July.
With 31 billion parameters, Solar Pro 2 still matched or even outscored bigger rivals on several tests. It was one of only two Korean LLMs to be listed on the UK-based Artificial Analysis index, and the only to be named a “frontier model.”
Placing first in “Intelligence vs. Cost to Run” — a measure of how much capability a model delivers for its operating cost — it prompted a fiery response from Elon Musk, who insisted his Grok 4 model “remains No. 1,” in response to a post by Artificial Intelligence hailing the Korean product.
The startup underscores the “exceptional” multilingual processing capabilities of Solar Pro 2, demonstrating a robust performance across a wide range of general Korean natural language processing. At the same time, Upstage is targeting knowledge-heavy fields like law, finance and health care, where efficiency and accuracy matter as much as raw computing power. As the only startup chosen for Korea’s national AI project, Upstage represents the growing role of nimble challengers in a space often dominated by chaebol and foreign tech firms.
With the Solar Pro 2 development experience as the foundation, Upstage aims to design a new architecture and training algorithm from scratch.
The anticipated model, Solar WBL, is targeting between 100 billion and 300 billion parameters with plans to evolve in stages — expanding in its language capabilities as well as multimodal functions and industry-specific versions.
South Korean internet giant Naver launched the country’s first major foray into LLMs with HyperClova in 2021. Four years later, its successor HyperClova X, developed by Naver Cloud, has emerged as the company’s flagship hyperscale model that excels in Korean language, culture and context, in comparison to other global peers.
Leveraging Naver’s massive Korean-language data set, HyperClova X has been trained using 6,500 times more Korean data than OpenAI’s GPT-4, according to the company.
Though Naver has not revealed the exact parameter count of HyperClova X, industry analysts believe it operates at a significantly smaller scale than major global models such as ChatGPT, Gemini or DeepSeek. Still, the model distinguishes itself in tasks in the Korean language.
HyperClova X outperforms GPT-4 in addressing Korean-specific questions, according to the KMMLU benchmark.
Under the HyperClova X umbrella, the lineup includes HyperClova X Think, a reasoning-specialized model, HyperClova X Dash, a lightweight variant, and HyperClova X Seed, a free, open-source version.
While HyperClova specializes in the Korean language, it boasts exceptional performance in Chinese, Japanese and Vietnamese as well.
Leveraging its expertise in Korean-language models, Naver, in partnering with US-based AI startup Twelve Labs and domestic universities, plans to build a proprietary omnifoundation model capable of processing diverse data types, including text, image, audio and video.
Kakao may not have made the cut for the five companies tasked with developing the country’s sovereign AI, but the internet giant behind messaging app KakaoTalk has been spearheading its own AI efforts using its large language model Kanana
Since debuting Kanana LLM last year, the company has steadily expanded the lineup: a lightweight Kanana Nano with 2.1 billion parameters, midtier Essence with 9.8 billion parameters and the powerhouse Flag with 32.5 billion parameters.
In May, the company pushed further into multimodal AI with Kanana-o, capable of processing text, image and speech, while also open-sourcing an upgraded Kanana 1.5.
The company’s AI ambition evolves around the Kanana brand. This spring, the company launched a standalone messenger app, also named Kanana, embedding an AI assistant directly into personal and group chats.
What’s drawing the most attention is Kakao’s much-anticipated tie-up with AI giant OpenAI, set to be unveiled in September. Though details are still largely under wraps, the two plan to roll out an AI agent by year’s end to operate across Kakao’s wide-spanning ecosystem of platforms.
Despite being a relative unknown in Korea’s AI sector, NC AI shocked the industry when it was selected for the government’s sovereign AI project.
Industry watchers attribute the initial success of NC AI, a subsidiary of gaming giant NCSoft, as deriving from a combination of 14 years of research and a consortium strategy bringing together academic, corporate and research partners.
NCSoft first established an AI unit in 2011, long before AI became a global trend, and accumulated AI technology across various areas.
In July 2023, it released Varco, becoming the first in Korea’s gaming industry to put out its own LLM. A year later, it introduced Varco 2.0, with 7 billion and 13 billion parameters for its small and base models, respectively, upgrading its capability.
More recently, the company last month unveiled Varco Vision 2.0, a multimodal model with 14 billion parameters for the full version and 1.7 billion parameters for the lightweight version, both in open source.
The model processes images and text simultaneously to answer questions. Despite its relatively smaller scale, the model outperformed other open-source counterparts, including InternVL3-14B, Ovis2-16B and Qwen2.5-VL 7B, across multiple benchmarks, from English and Korean image understanding to optical character recognition.
