How South Korea plans to best OpenAI, Google, others with homegrown AI 

From Tech Giants to Startups, South Korean players are developing large language models created for their own language and culture ready to compete with global heavyweights like OpenAI and Google.

Last monthThe country has launched its most ambitious sovereign AI initiative so far, promising five local companies to make a large -scale foundational model $ 530 billion, (about $ 390 million).

This step refers to Coul’s desire to strengthen national protection and reduce the dependence on foreign AI technology in hopes of reinforcing national protection in the AI ​​era.

The organizations pick up Ministry of Science and ICT Was competition LG AI research, SK Telecom, Navar cloud, NC AIAnd startup UpstageThe

Every six months, the government will review the progress of the first Kohar, cut the under -performers, and will continue to finance the frontiers until only two to lead the country’s sovereign AI drive.

Each player is bringing a different advantage in South Korea’s AI race. He talked about how Open, Google, Ethnic, and the rest of the parts were planning to take with several selected companies selected companies. NC AI refused to comment.

LG AI Research: Xon

LG AI Research, South Korean Giant LG Group Research and Development Unit Xon 4.0A hybrid argument AI model. The latest version is first introduced in the company’s previous Exon Deep model by combining broad language processing with advanced rational features.

Xon 4.0 (32b) has already scored reasonably well against competitors Artificial analysisIts detective index benchmark (eg like upstage) Solar pro 2) However, it is planned to improve and improve steps through deep access to real-world industries from biotech to advanced materials and production.

It is attaching the data to focus on the data refined before feeding in models for training. Instead of chasing the perfect scale, LG wants to make the entire process more intelligent, so that its AI can provide the original, practical value that goes beyond the general-principal models that can offer. “This is our basic method,” co-chief Hongklak Lee told TechCrunch.

LG is improving its models through familiar techniques: providing them through API, then using real-world data produced by the users of those services to train to improve the model.

“Our partners can provide better services as LG models are developed, resulting in greater economic value and even more prosperous data,” he said.

However, instead of chasing huge GPU clusters, LG AI is focusing on research skills, earning maximum from each chip and created art-specific models, he noted. The goal is not to overcome the global giants but to outmart with their high-performance, yet more efficient, AI.

South Korea’s Telco Giant SK Telecom (SKT) launched it Personal AI agent. (Pronunciation A-Dot) The service path returned to the end of the 2023, and this July only removed its new language model, ax.

Built at the top of the Chinese Open Source model from Alibaba Cloud, Qwen 2.5, X0 Two models, a huge 72-billion-parameter version and a light 7b version.

SK says that the X1.1 Korean GPT -4 and about 33% more efficiently input, under the native language edge. (OpenAIA GPT 5.0 comparison data is not available)) SKTO Open Sourced ITS The ax is 3.1 models At the beginning of this summer. Meanwhile, A. Service provides the AI ​​call summary and auto-exposed notes. It has already been drawn to about 10 million customers until August 2025.

The versatility of SK’s Edge is because it has access to information from its telecom network from navigation to taxi-hilling.

“The role of SK Telecom is to serve as a bridge between Cutting-Ez Model Research and Real-World effects. Our telecom infrastructure, with a broader user base and A. Served services like this, we bring AI AK directly to customer service, mobility or production,” Skilicom’s Foundation Foundation Model Taun Kim, said.

SK Telecom is also investing in South Korea’s largest GPU-based service GPUAS and is also investing in the AI ​​infrastructure by creating a new hypersiskel AI data center with AWS. Whatever the lack of it, it is being partner to get it.

“We are creating a full-scale ecosystem with the Korean AI Chipmaker uprising, protecting faithful data partnerships through work with government and universities,” Kim said, “Kim said,” Kim said. “These include projects like our cooperation with MIT (MGIC), which apply the Foundation models for advanced production and battery and semiconductor innovation.”

Navar Cloud: Hyperclava X

Cloud Services Bahu Navar Cloud, a leading Internet company in South Korea, introduced its large language model Hyperklova in 2021. Two years later, it unveiled an upgrade version, Hyperclava XWith new products powered by technology: Clova X, an AI chatbot and Q, a generator AI-powered search engine is located as a competitor to Microsoft’s Copilate-Bing and Google’s AI Overview. It has also unveiled the AI ​​model of this year’s multimodal argument, Hyperklov X thinksThe

Navar Cloud believes that the real power of the LLMS is a “connector” to serve as a connecting system and serving for the development of effectiveness as a sealed services, a spokesperson for NAV.

Navar stands as the only company in Korea – and one of the worlds in the world – which can really claim “II full stack”. It explains its hyperclova X model from scratch and operates huge data centers, cloud services, AI platforms, applications and consumer services that make the technology lively.

Like Google – however, for South Korea – Navar is embedded in main services such as searching, shopping, maps and finances in its AI. The benefit of this is real-world data. This is the AI ​​Shopping Guide, for example, people provide recommendations based on what they really want to buy. Other services include Clova Studios, which allows business to create custom generator AIs, and Clova Kerakal, an AI-driven check-in service prepared for seniors living alone.

Navar’s spokesperson says that Global AI giants like OpenAI and Google depend on two things: to make its “recipe” for models and secure capital to scale them. Nevertheless, instead of chasing the size, the company emphasizes sophistication, arguing that its AI is already competitive on comparative scales worldwide.

Upstage solar pro 2

The only startup upstage to compete in the project. Its solar Pro 2 Model, was launched last July, it was the first Korean model recognized as a border model Artificial analysisAccording to Open-El Cowon, Open AAEA, according to the Executive Vice President of the Openstage, is ringing it with Google, Meta and anthropologists.

Quoan told TechCrunch that most border models have 100 billion to 200 billion parameters, solar pro does 2-21 billion co-south well-performing better for Koreans and told Quaan TechCrunch more expensive.

“Solar Pro2 has surpassed Global models in Korean criteria. With this project, upstage aims to achieve 105% Korean language performance worldwide,” said Quan.

He said that upstage’s goal is to distinguish yourself by focusing on true business influence, not just criteria, he said. So it is developing special models for industries such as money, law and drugs while pressing for the creation of the Korean AI ecosystem under the leadership of “AI-Native” startups.