Details on K-Exaone AI Model

LG AI Research has recently disclosed the technical details of K-Exaone, developed as part of Korea’s initiative for a “homegrown AI foundation model.” The announcement highlights the model’s basis on an independent technical design, showcasing Korea’s efforts to define what qualifies as homegrown AI amid ongoing debates in the ICT industry.

K-Exaone has made significant strides, being recognized as the only Korean-developed AI model to secure a position in the top 10 of a global ranking by an AI benchmarking organization. This benchmark acknowledgment emphasizes its competitive performance on an international scale.

On January 11, LG AI Research released a technical report that outlines K-Exaone’s architecture design, training methods, and comprehensive evaluation results. The model was developed in-house over the past five years, incorporating advancements in “hybrid attention,” a critical technology first introduced in Exaone 4.0. This innovation has enabled LG to reduce memory and computational requirements by 70% compared to earlier versions.

Such improvements in design not only enhance efficiency but significantly lower operational costs, allowing K-Exaone to function effectively in an A100-class GPU environment without the need for costly, high-end infrastructure. This accessibility is a boon for companies operating with limited resources, thereby broadening the base of Korea’s AI ecosystem.

In terms of performance, K-Exaone ranked seventh globally and first among Korean models in a leaderboard compiled by the AI evaluation group Artificial Analysis. The list predominantly showcases six Chinese models and three from the United States, with K-Exaone standing out as the sole Korean entry. The model’s application programming interface (API) will be provided free of charge on the institute’s website until January 28.

LG AI Research’s Choi Jung-kyu, who leads the agentic AI division, expressed the significance of K-Exaone by stating that it demonstrates how a model based on an independent technical design can effectively compete with global models despite limited resources. The institute is focused on developing models that not only serve local needs but also contribute positively to the broader global AI landscape.

Further emphasizing its robust technological foundation, K-Exaone achieved an average score of 72 points across 13 evaluations during the first round for the government’s homegrown AI foundation model project. This score placed it at the forefront compared to competitors, which included Naver Cloud, Upstage, SK Telecom, and NC AI. The initiative, spearheaded by the Ministry of Science and ICT, aims to create an AI model that ranks among the top 10 globally, utilizing domestic technology for public sector applications.

However, the initiative has not been without controversy, as questions have been raised concerning the homegrown status of some participants. For instance, Naver Cloud faced scrutiny for allegedly utilizing open-source technology from Alibaba, while Upstage’s similarities to China’s Zhipu AI and SK Telecom’s resemblance to DeepSeek were also highlighted. The Ministry is expected to conclude the first round of evaluations soon, with one team poised for elimination.

This reporting was originally drafted in Korean and has undergone translation via generative AI tools. It was subsequently refined by an editor with native English proficiency, ensuring accuracy in the presentation of information about K-Exaone.