LG AI Research released the K-EXAONE 2.0 model on Hugging Face today local time. As its name suggests, this is the second model developed by LG AI Research under the Sovereign AI Foundation Model Project of South Korea's Ministry of Science and ICT.
K-EXAONE 2.0 uses a hybrid-attention MoE architecture, with 750B total parameters and 37B active parameters. It is more than three times the size of the first-generation K-EXAONE, which had 236B-A23B parameters, and is the largest AI foundation model developed in South Korea to date. The new model is open-sourced under the Apache 2.0 license.
In a longitudinal comparison, K-EXAONE 2.0 achieved an average score of 70.1 across 24 benchmarks, more than 10% higher than the first-generation K-EXAONE's score of 63.3. Its performance improved by approximately 30% on the three major coding and agentic coding benchmarks.
In a cross-model comparison, the model outperformed GLM-5.1 in long-context understanding and surpassed Zhipu GLM-5.1 and Qwen3.5 in agentic tool-use capabilities. In instruction-following evaluations, it demonstrated performance comparable to GLM-5.1, DeepSeek V4 Pro Max, Qwen3.5, and others.