LG AI Research Releases K-EXAONE 2.0, a 750B Parameter Open-Weights Model
LG AI Research has released K-EXAONE 2.0, a 750-billion parameter (37 billion active) Mixture of Experts (MoE) open-weights model under the Apache 2.0 license. The model was developed as part of South Korea's government-funded sovereign AI initiative. This release marks a significant milestone in South Korea's national AI project, which aims to build domestic AI capabilities and reduce reliance on foreign technology providers. By offering the model under an open-source license, it provides the global AI community with a massive, highly capable MoE model. The model is available in multiple quantized formats, including FP8 and NVFP4, the latter designed for efficient low-precision inference on NVIDIA Blackwell GPUs. The project is part of a competitive government program where participating companies are periodically evaluated and filtered out.
## BACKGROUND
Sovereign AI refers to national or regional efforts to develop independent AI infrastructure, models, and data controls to protect national security and cultural representation. Mixture of Experts (MoE) is an architecture that activates only a subset of a model's parameters (experts) for any given input, drastically reducing computational costs during inference. NVFP4 is a 4-bit floating-point format introduced by NVIDIA to maintain model accuracy while reducing memory bandwidth and storage requirements.