Lee Gyu-lee is a business writer at The Korea Times, focusing primarily on IT & telecommunications, the Ministry of Trade, Industry and Energy and KOTRA. Prior to this, she has covered a wide range of cultural news, from film, television and K-pop to lifestyle and fashion.
LG's K-EXAONE delivers top benchmark results
A graph comparing five national AI foundation project models / Courtesy of LG AI Research
By Lee Gyu-lee
LG AI Research’s newly developed artificial intelligence (AI) model, K-EXAONE, created for the national AI foundation model project, has demonstrated strong competitiveness in both domestic and international AI arenas.
The company shared that K-EXAONE topped 10 out of 13 benchmark tests in the first-round evaluation of the government-backed AI foundation model initiative, scheduled to conclude this month.
The model earned an average score of 72 points, delivering the strongest overall performance among models developed by five consortia participating in the project.
K-EXAONE scored 32 points on the Artificial Analysis Intelligence Index, ranking first in Korea and seventh worldwide among open-weight models, according to the global AI analysis platform.
The Korean model was also the only non-Chinese or U.S. entrant in the top 10, which was dominated by six Chinese and three U.S. models.
“We planned the model’s development under realistic timelines and infrastructure constraints, using only about half of our available training data to build K-EXAONE,” said Lee Jin-sik, leader of LG AI Research’s EXAONE Lab.
“This initial release is a starting point toward a frontier-grade model, and we plan to significantly scale up performance from this point onward.”
After releasing its open weights on the Hugging Face, K-EXAONE climbed to No. 2 on the platform’s global model trend ranking, drawing strong interest from AI researchers worldwide.
The model was also recognized by the U.S. nonprofit AI research institute Epoch AI for its Notable AI Models list. With this, LG AI Research has become the first Korean company to place five models on the list, including EXAONE Deep and EXAONE 4.0.
The list is widely referenced in Stanford University’s annual AI reports and is used as a benchmark for assessing national and corporate AI competitiveness.
K-EXAONE represents the five years of LG AI Research’s in-house foundation model R&D, integrating its proprietary technologies to achieve higher efficiency at lower training and operational costs. Rather than simply scaling by data size, the team re-engineered the model architecture itself.
The model incorporates an advanced version of hybrid attention that determines which information an AI model should focus on when processing massive datasets.
LG AI Research refined its hybrid attention mechanism by combining sliding-window attention for local focus with global attention for full-context understanding. This approach cuts memory use and computational load by 70 percent compared to the previous model, Exaone 4.0.
K-EXAONE has a total parameter count of 236 billion and adopts a mixture-of-experts architecture, with only about 23 billion parameters activated at a time. Parameters are the internal variables that an AI model learns from data during training to make predictions.
“K-Exaone’s efficient architecture delivers frontier-level performance even on A100-class graphics processing units, rather than depending on ultra-expensive infrastructure. This design will allow resource-limited organizations to deploy powerful large models, helping expand the domestic AI ecosystem,” an LG official said.