LG AI Research Institute Releases Korea's Largest 75 Billion Parameter 'K-EXAONE 2.0'

Second Phase of the 'Independent AI Foundation Model'

Commercial Use Enabled Under 'Apache 2.0' License

Outperforming Leading Global Models

On July 31, LG AI Research Institute announced that it has open-sourced the second phase AI model of the 'Independent AI Foundation Model Project', led by the Ministry of Science and ICT, by releasing 'K-EXAONE 2.0' to the Hugging Face open-source community.

LG AI Research Institute has released 'K-EXAONE 2.0', which has 75 billion (750B) parameters—the largest in Korea—and applied the 'Apache 2.0' license that allows commercial use. This model has more than tripled in parameter size compared to the first-phase model (236B).

'K-EXAONE 2.0' achieved an average score of 70.1 across 24 benchmark evaluations in nine performance categories, surpassing the first phase’s (63.3 points) performance by more than 10%. In particular, the average score for key indicators in coding and agentic coding areas surged by 30%.

In the long-context comprehension areas, 'K-EXAONE 2.0' scored 94.4 points on the global long-context understanding benchmark (OpenAI-MRCR) and 89.6 on the Korean context understanding evaluation (Ko-LongBench), significantly outperforming Zhipu's GLM-5.1, which scored 71.5 and 83.6, respectively. The model also demonstrated performance on par with leading global models such as GLM-5.1, DeepSeek V4 Pro max, and Qwen 3.5 in agentic tool use ability (Tau3-Bench Banking, 14.2 points) and instruction following evaluations (IFEval, IFBench).

LG AI Research Institute has expanded the supported languages of 'K-EXAONE 2.0' from six to ten by adding French, Italian, Portuguese, and Polish, thereby establishing the most multi-lingual sovereign AI framework among domestic AI foundation models. In safety evaluations reflecting both Korea’s specific context and global ethical standards (KGC-Safety, ROK-Fortress), the model achieved an average score of 94.6, proving its technical safety.

Earlier, in April, LG AI Research Institute released the vision-language model (VLM) ‘EXAONE 4.5’, which has since been deployed in real-world settings such as the Ministry of the Interior and Safety's 'AI Safety Report' and the Ministry of Food and Drug Safety's 'AI Drug Review' system. ‘EXAONE 4.5’ outperformed GPT5-mini, Claude Sonnet 4.5, and Qwen3-VL in the average of 13 benchmarking indicators, including three for general visual comprehension, infographics combining images and text, and five document understanding and reasoning metrics for complex information in specialized literature.

The institute plans to further strengthen its 'expert AI' portfolio by releasing industry-specialized AI foundation models next week, aiming to drive innovation in industry and daily life. Additionally, a new service will be launched to allow everyone in Korea to experience 'K-EXAONE 2.0'.

This content was produced with the assistance of AI translation services.