Home > Business > Tech

LG AI Research discloses details of K-Exaone, developed for 'homegrown AI foundation model' initiative

Published: 11 Jan. 2026, 18:40

Attendees try out LG AI Research’s technology at the first presentation of the independent AI foundation model project at Coex in Seoul’s Gangnam District, southern Seoul, in December 2025. [YONHAP]

LG AI Research has disclosed technical details of K-Exaone, which it developed for Korea's “homegrown AI foundation model” initiative, emphasizing that it is built on an independent technical design as the ICT industry continues to debate what qualifies as homegrown AI.

The company said the foundational model is the only Korean-developed AI model to have entered the top 10 in a global ranking by an AI benchmarking organization, meaning its performance has been recognized internationally.

LG AI Research said on Sunday that it released a technical report detailing K-Exaone's architecture design, training methods and evaluation results.

The company developed K-Exaone in-house over the past five years. By further advancing “hybrid attention,” a core technology introduced in Exaone 4.0, LG AI Research reduced memory requirements and computation by 70 percent compared to previous versions.

“The model’s design improves efficiency while lowering costs, allowing it to run in an A100-class GPU environment rather than requiring expensive, high-end infrastructure,” said an LG AI Research representative. “This could help broaden the base of Korea’s AI ecosystem by making it usable for companies with limited infrastructure.”

Attendees look around the Naver Cloud exhibition area at the first presentation of the independent AI foundation model project at Coex in Seoul’s Gangnam District, southern Seoul, in December 2025. [YONHAP]

K-Exaone ranked seventh worldwide and first in Korea in an global open weight AI leaderboard compiled by the AI evaluation group Artificial Analysis. Open weight models are those whose weights are publicly available for download, use and modification, though key details such as training data remain undisclosed.

With six Chinese models and three U.S. models dominating the top 10, K-Exaone is the only Korean model to make the list. The K-Exaone application programming interface, or API, will be available for free on the institute’s website through Jan. 28.

“K-Exaone shows that a model based on an independent technical design can compete on par with global large-scale models despite resource constraints,” said Choi Jung-kyu, the head of the agentic AI division at LG AI Research. “The institute aims to build models that contribute to the global AI ecosystem beyond Korea.”

Attendees look around the SK Telecom exhibition hall at the first presentation of the independent AI foundation model project at Coex in Seoul’s Gangnam District, southern Seoul, in December 2025. [YONHAP]

LG also highlighted that K-Exaone scored an average of 72 points across 13 tests during the first round of evaluations for the government's homegrown AI foundation model project, the highest among the five candidates participating in the initiative: Naver Cloud, Upstage, SK Telecom, NC AI and LG AI Research.

Led by the Ministry of Science and ICT, the project aims to develop an AI model with top 10 global performance using domestic technology and deploy it for use by public institutions and others.

However, controversy has emerged over whether some participants qualify as “homegrown.” Naver Cloud was later found to have used open-source technology from China’s Alibaba. Claims have also surfaced that Upstage resembles China’s Zhipu AI, and that SK Telecom’s configuration is similar to that of China’s DeepSeek.

The ministry is set to complete the first round of evaluations by Thursday and announce one team to be eliminated.

This article was originally written in Korean and translated by a bilingual reporter with the help of generative AI tools. It was then edited by a native English-speaking editor. All AI-assisted translations are reviewed and refined by our newsroom.
BY KIM KYUNG-MI [[email protected]]