South Korea's AI Models: A Guide to Homegrown LLMs

South Korea has become one of the few countries with a serious domestic ecosystem of large language models. Driven by a government push for "sovereign AI," strong chip and cloud industries, and the need for models that understand Korean language and culture, companies such as Naver, LG, SK Telecom, Kakao, KT, Upstage and NC AI have each built their own model families. Many are released with open weights, and most are tuned for Korean first and English second. This article lists the main models, then breaks down the features and strengths of each. Model lineups change quickly, so check each developer's site for the latest versions.

Overview Table

HyperCLOVA X (Naver)

Naver's HyperCLOVA X is among the earliest and best-known Korean foundation models. It powers Naver's search and assistant products and is offered to businesses through Naver Cloud.

Features

• Trained on large volumes of Korean-language data
• Multimodal capability, including image understanding
• Available through cloud APIs, with smaller open-weight variants released for developers

Strengths

• Excellent grasp of Korean idiom, honorifics and local context
• Tight integration with Naver's search, shopping and maps ecosystem

EXAONE (LG AI Research)

EXAONE comes from LG's research arm and targets professional and industrial use, including science, coding and mathematics.

Features

• Multiple sizes, from small on-device models to large reasoning models
• Reasoning-focused variants for math, coding and scientific tasks
• Bilingual Korean and English performance, with open-weight releases for research

Strengths

• Competitive scores against global models of similar size
• Strong fit for enterprise, manufacturing and research workflows

Solar (Upstage)

Upstage is a start-up that gained attention by showing that compact models can rival much larger ones. Its Solar family is built around efficiency.

Features

• Compact architectures designed for lower compute cost
• Document-processing tools, including document parsing and OCR products
• API access and options for private deployment

Strengths

• Strong performance per parameter, which lowers serving costs
• Practical focus on business documents and retrieval workflows

A.X (SK Telecom)

SK Telecom's A.X series is part of the company's strategy to become an AI-centered enterprise, building on its earlier A. assistant services.

Features

• Korean-optimised language models for enterprise and telecom services
• Efforts toward large-scale models under national AI programmes
• Integration with SK's network, data-centre and customer-service operations

Strengths

• Backing from a major telecom and infrastructure operator
• Strong potential for call-centre, customer-support and enterprise automation

Kanana (Kakao)

Kakao, the company behind the KakaoTalk messenger, designed Kanana for products used by millions of people every day.

Features

• A range of model sizes, including lightweight versions
• Language, multimodal and embedding models
• Open-weight releases for developers and researchers

Strengths

• Practical, cost-efficient models suited to chat and consumer apps
• Natural conversational tone in Korean

Mi:dm (KT)

KT's Mi:dm (short for Mutual Intelligence) is positioned for business and public-sector customers who need Korean-language accuracy and local data control.

Features

• Korean-centric training with English support
• Variants designed for enterprise and government deployment
• Open-weight releases of smaller models

Strengths

• Emphasis on data sovereignty and domestic hosting
• Established relationships with Korean corporate and public clients

VARCO (NC AI)

NC AI spun out of the game developer NCSOFT, and VARCO reflects that heritage with models aimed at creative and entertainment work.

Features

• Language models alongside generative image, voice and 3D tools
• Tools for game development, character dialogue and content creation
• Korean and English language support

Strengths

• Domain expertise in games, media and digital humans
• Creative-production focus that few other Korean models share

Other Notable Efforts

Samsung's Gauss models are used mainly inside the company and on devices, while start-ups such as Motif Technologies and Trillion Labs are building efficient open models. The government's sovereign AI foundation model programme has also been narrowing its list of participating teams, so the competitive picture continues to shift.

Korea's AI landscape is diverse. Naver and Kakao lean toward consumer products, LG and Upstage toward reasoning and efficiency, SK Telecom and KT toward telecom and enterprise, and NC AI toward creative industries. Their shared advantage is native fluency in Korean, which global models often handle less naturally. For anyone choosing among them, the right pick depends on the use case: pick Naver for local knowledge, LG for reasoning, Upstage for cost efficiency, or the telecom players for enterprise integration.

› More Article