LG AI Research Institute officially released the K-EXAONE 2.0 model on Hugging Face on July 31, which is the second-generation product developed under the Sovereign AI Foundation Model Project of the Korea Ministry of Science and Information and Communication Technology. The new model adopts a hybrid attention MoE architecture, with a total of 750 billion parameters and 37 billion activated parameters, more than three times the size of the first generation, and is also the largest artificial intelligence foundation model ever developed in South Korea.
Notably, K-EXAONE 2.0 is fully open source under the Apache 2.0 license, which is quite rare in the history of AI development in South Korea.
Performance has been significantly improved, directly competing with Chinese open-source models
Vertically, K-EXAONE 2.0 scored an average of 70.1 points in 24 benchmark tests, an increase of over 10% compared to the initial version's 63.3 points. Performance improvements reached about 30% in three coding and agent coding benchmark tests. Horizontally, it is even more ambitious - it outperforms Zhipu GLM-5.1 in long context understanding, and it also surpasses GLM-5.1 and Qwen3.5 in agent tool usage capabilities. In instruction execution evaluation, it is on par with GLM-5.1, DeepSeek V4 Pro Max, and Qwen3.5.
From the first generation to 2.0, LG AI Research Institute increased the model parameter scale by more than three times within one generation, and directly competes with top Chinese open-source models in key indicators. With the global competition for open-source models triggered by Kimi K3 still heating up, South Korea has joined the race under the name of sovereign AI, indicating that the triangular competition pattern in the East Asian open-source AI ecosystem is taking shape.
