From technological giants to newly established businesses, South Korean players develop large linguistic models adapted to their own language and culture, ready to compete with worldwide heavyness such as Openai and Google.
Last monthThe nation has so far began the most ambitious AI initiative, blocking $ 530 billion (about $ 390 million), to five local companies that build large -scale basic models.
The move highlights Seoul’s desire to reduce the dependence on foreign AI technologies, hoping to boost national security and maintain stricter data control in the AI era.
The organizations who chose the Ministry of Science and ICT to compete were Research lg ai; SK telecommunication; Cloud; Nc aiand starting Upstage.
Every six months, the government will revise the progress of the first coil, reduce the degradable and continue to fund the pioneers until only two remain to lead the country’s dominant driving.
Each player brings a different advantage to the South Korean AI match. TechCrunch has spoken to several of the selected companies about how they plan to take Openai, Google, Anthropic and the rest in the lawn of their home. NC AI refused to comment.
LG AI research: exaone
LG AI Research, South Korea’s Giant LG Giant LG unit, offers Exaone 4.0A hybrid logical model AI. The latest edition combines wide language processing with advanced reasoning features that were first introduced into the company’s previous Exaone Deep model.
TechCrunch event
Francisco
|
27-29 October 2025
Exaone 4.0 (32B) is already rating pretty well against competitors Artificial analysisInformation Index (as well as upstage’s upstage’s Solar pro2). But she plans to improve and climb classes through her deep access to real -world data ranging from biotechnologies to advanced materials and manufacturer.
It connects these data with emphasis on refinement of data prior to diet with models to train. Instead of chasing a clear scale, LG wants to make the whole process smarter so that AI can provide a real, practical value that transcends general -use models. “This is our fundamental approach,” Honglak Lee said, Honglak Lee coincides.
LG improves its models through well -known tactics: offering them through API and then using the real world data produced by users of these services to train the model to improve.
“As LG models improve, our partners can offer better services, which in turn create greater economic value and even richer data,” he said.
However, instead of hunting for massive GPU clusters, LG AI research focuses on efficiency, making the most of every chip and creating models that specifically for industry, he said. The goal is not to overcome world giants, but to overcome them with high performance, but more effective, AI.
South Korea’s Sk Telecom (SKT) giant of South Korea started Personal agent ai a. (A-DOT) is pronounced back in the late 2023 and just started the new big language model, AX, in July.
Built on top of the Chinese open source model by Alibaba Cloud, Qwen 2.5; AX 4.0 It comes in two models, an intense version of 72 billion parameters and a lighter 7b version.
SK says that AX 4.0 processes the Korean inflows about 33% more efficient than the GPT-4O, underlining the edge of the local language. (Openai’s GPT 50 comparison data is not available.) SKT is also open AX 3.1 models Earlier this summer. In the meantime, A. A. offers features such as AI call summaries and notes that are automatically created. Since August 2025, it has already drawn about 10 million subscribers.
The tip of SK is its flexibility because it has access to information from the telecommunications network ranging from navigation to sorting.
“The role of SK Telecom is to act as a bridge between cutting -edge models and real effect.
SK Telecom also invests in AI infrastructure, using GPUAAS, the largest South Korean GPU -based service and the building of a new AI AI data center. Whatever it is not, he works to get to acquire.
“We are creating a complete ecosystem with Korean AI Chipmaker uprisings, securing reliable corporate data relationships through work with government and universities and promoting a global research network,” Kim said. “This includes projects such as our collaboration with MIT (MGAIC), which applies foundation models to advanced construction and battery and semiconductor innovations.”
Naver Cloud: Hyperclova x
Naver Cloud, the South Korean Cloud Services arm, introduced its large linguistic model, HyperClova, in 2021. Two years later, it presented an upgraded version, Copilot-Enhiling and Generation Ai-Driven Search Points at Poirs Ai-Driven Search. He also presented this year’s multimodal logical model AI, Hyperclove x think.
The Naver Cloud believes that the real power of LLMS is to serve as “fasteners” that connect bequeathed systems and Siled services to improve utility, according to a Naver spokesman.
Naver stands out as Korea’s only company – and one of the few in the world – who can really claim to have a “full stack AI”. He created the Hyperclova X model from scratch and performs huge data centers, cloud services, AI platforms, applications and consumer services that bring technology to life, the spokesman explained.
Similar to Google – but coordinated for South Korea – Naver incorporates AI into basic services such as search, shopping, maps and funding. Its advantage is real world. It is the AI shopping guide, for example, it offers recommendations based on what people really want to buy. Other services include Clova Studio, which allows businesses to create customized AI and Clova Carecall, a check-in AI-powered check-in service.
Naver’s spokesman says that Global Ai Giants like Openai and Google depends on two things: perfecting his “recipe” for models and securing the capital to escalate. However, instead of chasing the size, the company emphasizes complexity, arguing that its AI is already competitive worldwide on comparable scales.
The upstage Solar Pro 2
Upstage is the only startup that competes in the project. The Solar Pro 2 model, launched last July, was the first Korean model recognized as a Frontier model by Artificial analysisPutting it in the ring with Openai, Google, Meta and Anthropic, according to the Soon-Il Kwon, an executive vice president of the upstage.
While most Frontier models have 100 billion to 200 billion parameters, the Solar Pro 2-with just 31 billion-performs better for the southern Koreans and is more cost-effective, Kwon in Techcrunch.
“Solar Pro 2 has surpassed world models at important Korean reference points. With this project, upstage aims to achieve a Korean linguistic performance of 105% of the world model,” Kwon said.
The upstage aims to diversify by focusing on real business impacts, not just on benchmarks, he said. It therefore develops specialized models for industries such as funding, law and medicine, while pushing to build an AI Korean ecosystem led by “Ai-Native” startups.
