K-EXAONE is a large-scale multilingual language model developed by LG AI Research as part of South Korea's national AI foundation model project. It is architecturally distinct from the earlier dense EXAONE series, adopting a Mixture-of-Experts (MoE) structure designed for efficient scaling and advanced logical reasoning. The model is optimized for complex tasks in mathematics, coding, and science, utilizing synthesized data including "thinking trajectories" to enhance its internal reasoning capabilities.
Technically, K-EXAONE features a 236B total parameter count with 23B active parameters during inference. It incorporates a Multi-Token Prediction (MTP) module that enables self-speculative decoding, reportedly increasing inference speed by approximately 1.5 times compared to standard autoregressive models. The architecture also employs a 3:1 hybrid attention scheme, combining sliding-window and global attention to minimize memory usage during long-document processing.
The model supports a native context window of 256,000 tokens and provides multilingual coverage for six languages: Korean, English, Spanish, German, Japanese, and Vietnamese. Its vocabulary is expanded to 150,000 tokens using a proprietary SuperBPE method to improve tokenization efficiency. K-EXAONE is trained to align with human preferences through a specialized preference learning phase called GrouPER (Group-wise SimPER) and is designed to respect Korean cultural and historical contexts alongside universal safety standards.