As of Dec. 2025, EXAONE (LG AI Research) 4.0 32B (Reasoning) secured about 82% of the performance compared with U.S. GPT-5, narrowing the technology time gap to 5.9 months.
Choi Jeong-gyu, head of the AI Agent Group at LG AI Research Institute, stated accordingly at the first briefing of the independent AI foundation model project held Dec. 30 at the COEX Auditorium in Samseong-dong, Seoul. Choi said, "Compared with DeepSeek, cited as China's representative model, we are at a comparable level with a 0.6-month gap, and we outperformed representative models from France, the United Kingdom, and Canada," adding, "We achieved a learning efficiency of 89.4%, and even at the 32B scale, we demonstrated performance on par with top-tier models in the hundreds of billions to 1T range." A step-by-step learning strategy moving from general knowledge to advanced knowledge and specialized domain knowledge was also presented as a differentiator.
The performance indicators focused on the "three types of reasoning." Choi explained, "We recorded 92.9% in mathematics (AIME 2025), 80.4% in scientific reasoning (GPQA Diamond), and 81.2% in coding (LiveCodeBench), delivering results comparable to global SOTA models in high-difficulty areas." Choi added, "In 'Tau2 Bench Telecom,' which shows practical tool-using capabilities, we also recorded 68.2%, increasing the potential for agentic AI adoption."
Choi also expressed an ambition to usher in an era of global top 3 with "K-EXAONE." The aim of K-EXAONE is the "expansion of on-site application" of research achievements. Choi introduced, "We enhanced cost competitiveness by automating scheduling from naphtha feedstock receipt to cracking in petrochemical processes, and through EXAONE (LG AI Research) Discovery, we shortened the discovery of cosmetic ingredients from 22 months to one day."
Choi added, "For enterprise use, we are boosting productivity with ChatEXAONE, which supports search, document summarization, and coding, and we are also providing customized AI platforms to the Gyeonggi Provincial Office of Education, the Korean National Police Agency, and NH NongHyup Bank in the public institutional sector." In the business-to-consumer (B2C) area, Choi explained that the reach has expanded by integrating into LG Uplus' AI call app "ixi-O," providing voice phishing detection and call summarization functions.
LG AI Research Institute formed the K-EXAONE consortium with Hancom, LIG, DOUZONE BIZON, and FuriosaAI to broaden the value chain and unveiled a plan to provide a high-performance on-premises environment as an integrated package with the domestic AI accelerator product (RNGD). The strategy is to build a platform as a service (PaaS) environment that allows anyone to easily develop AI agents and to spread AX (AI transformation) across industry.
Choi said, "By 2027, we will solidify our status as a frontier (pioneer) model through a four-stage roadmap and complete a Korea-style sovereign AI ecosystem by bundling infrastructure, accelerators, platforms, and services," adding, "K-EXAONE will be the centripetal force that elevates the Republic of Korea to the global AI top three."