LG Unveils Proprietary AI Foundation Model "K-ExaOne," Ranks 7th Globally

Ranked First Among Five Government Project Teams
Entered Global AI Performance Rankings
Five Years of Research, Concentrated Technological Expertise

LG AI Research Institute has unveiled "K-ExaOne (EXAONE)," its proprietary AI foundation model developed by integrating its own technologies. The model ranked seventh in the global top 10 AI (artificial intelligence) models-a list otherwise dominated by the United States and China-making it the only non-U.S. or Chinese entry.

On January 11, LG AI Research Institute announced that K-ExaOne achieved an overall average score of 72 points across 13 benchmark tests, the primary evaluation criteria for the government's independent AI foundation model project, securing first place among five elite teams. K-ExaOne ranked first in 10 out of the 13 tests.

K-ExaOne also scored 32 points in the intelligence index assessment by Artificial Analysis, a global AI performance evaluation agency, placing it seventh worldwide. Among the top 10 open-weight models, six are from China and three from the United States, with K-ExaOne being the only other entry.

Immediately after its release, K-ExaOne ranked second in the global model trend rankings on Hugging Face, the global open-source AI platform.

K-ExaOne Model Architecture, LG Electronics

Additionally, it was included in the "Notable AI Models" list by Epoch AI, a U.S.-based nonprofit AI research organization. With this, LG AI Research Institute became the Korean company with the most entries-five models-on the list in 2024, including "ExaOne 3.5," "ExaOne Deep," "ExaOne Pass 2.0," "ExaOne 4.0," and "K-ExaOne."

LG AI Research Institute designed K-ExaOne for "high efficiency and low cost," enabling it to run on NVIDIA A100-class graphics processing units (GPUs) rather than requiring expensive infrastructure.

The institute further advanced "hybrid attention" technology, reducing memory and computational requirements by 70% compared to ExaOne 4.0.

K-ExaOne's training vocabulary consists of 150,000 words. Through the advancement of its "tokenizer" technology-which splits sentences into tokens, the units understood by AI-the model can remember and process documents 1.3 times longer than previous models. Its inference speed has improved by 150% over earlier versions.

Moreover, among domestic AI models, K-ExaOne can process the longest context-260,000 tokens-in a single pass, which is equivalent to over 400 pages of A4 documents.

LG AI Research Institute has verified the safety and reliability of K-ExaOne through its in-house AI Ethics Committee. The institute conducts data compliance evaluations for all training data, identifying and excluding data with potential copyright issues in advance.

K-ExaOne scored an average of 97.83 points across four categories on the "KGC-SAFETY" index, which LG AI Research Institute developed to assess Korean-specific characteristics. This is higher than the 92.48 points of OpenAI's GPT-OSS 120B model from the United States and the 66.15 points of Alibaba's Qwen3 235B model from China.

Choi Jeongkyu, Head of the Agentic AI Group at LG AI Research Institute, said, "K-ExaOne is a case that demonstrates how proprietary technical design can enable competition with global large-scale models, even with limited resources. With confidence in developing Korea's representative AI, we will focus on research and development to create a model that contributes to the advancement of the global AI ecosystem."