LG AI Talk Concert 2025 (Image courtesy of Yonhap)
SEOUL, Jan. 12 (Korea Bizwire) — LG AI Research on Sunday unveiled K-Exaone, a homegrown artificial intelligence foundation model that has placed South Korea among the world’s top-tier AI developers, according to global benchmark results and independent evaluations.
The model ranked seventh worldwide on the Intelligence Index compiled by Artificial Analysis, a global AI performance evaluator, making it the only non-U.S. and non-Chinese system to break into the top 10. Of the open-weight models ranked, six were from China and three from the United States.
LG AI Research said K-Exaone also topped the government’s first-round evaluation for South Korea’s independent AI foundation model project, scoring an average of 72 points across 13 benchmark tests and ranking first in 10 of them among five competing elite teams.
Following its release, K-Exaone quickly climbed to second place on Hugging Face’s global model trend rankings, reflecting strong interest from the open-source AI community.
It was also named a “notable AI model” by Epoch AI, a U.S.-based nonprofit research institute. With this addition, LG AI Research now has five models recognized by Epoch AI, the most among South Korean companies.
The institute emphasized K-Exaone’s efficiency-driven design, noting that the model can operate on Nvidia A100-class graphics processing units rather than requiring cutting-edge, high-cost infrastructure.
By advancing its hybrid attention architecture, LG reduced memory and computation demands by about 70 percent compared with its previous Exaone 4.0 model, while improving inference speed by 150 percent.
LG AI Research Launches EXAONE 4.0, Pioneering Hybrid AI in South Korea in 2025 (Image supported by ChatGPT)
K-Exaone supports a vocabulary of 150,000 tokens and can process up to 260,000 tokens of context in a single pass, equivalent to more than 400 pages of A4 text — the longest context length among domestic AI models, LG said.
The company also highlighted its emphasis on safety and compliance. All training data underwent copyright screening, and the model was evaluated by LG’s internal AI ethics committee. On LG’s Korea-specific safety benchmark, KGC-SAFETY, K-Exaone scored 97.83 points, outperforming comparable models from OpenAI and Alibaba, according to the institute.
“K-Exaone demonstrates that with independent design and efficient engineering, it is possible to compete head-to-head with global mega-models despite limited resources,” said Choi Jung-kyu, head of agentic AI at LG AI Research. “Our goal is to build a representative Korean AI that contributes meaningfully to the global AI ecosystem.”
The release comes as South Korea intensifies efforts to establish itself as a global AI powerhouse alongside the United States and China, positioning domestic foundation models as a strategic pillar of national competitiveness.
Kevin Lee (kevinlee@koreabizwire.com)
