LG AI Research said on the 31st that it released the second model of the Ministry of Science and ICT's "independent artificial intelligence (AI) foundation model project," "K-EXAONE 2.0," to the open-source community Hugging Face.
K-EXAONE 2.0 has 750B (75 billion) parameters, the largest in Korea, more than tripling the first model (236B). LG AI Research fully opened the model license policy to allow anyone to use it commercially.
K-EXAONE 2.0 scored an average of 70.1 on 24 benchmark evaluation metrics, achieving performance more than 10% higher than the first model (63.3). In particular, the average performance of three key indicators in coding and agentic coding rose 30%.
K-EXAONE 2.0's supported languages expanded from Korean, English, Spanish, German, Japanese, and Vietnamese to include French, Italian, Portuguese, and Polish, for a total of 10. This is the highest among AI foundation models in Korea.
K-EXAONE 2.0 also achieved an average score of 94.6 on the assessment of compliance with Korea's particularities and global ethical standards (KGC-Safety) and the safety evaluation related to geopolitical context judgment (ROK-Fortress), scoring higher than global leading models.
Im Woo-hyeong, co-head of research at LG AI Research, said, "K-EXAONE 2.0 is meaningful in that domestic researchers independently completed the entire process of a 750B-class model, from design and data training to distributed training and inference environment setup," and emphasized, "We have secured the capability to compete in the same weight class as global frontier models."
LG AI Research plans to further unveil a new industry-specialized AI foundation model next week, strengthening its "expert AI" portfolio.
During the second review of the "independent AI foundation model project," the development of a public evaluation site for practical usability assessment is currently being finalized, and a service is being prepared to allow all citizens to experience "K-EXAONE 2.0."
