EXAONE 4.0 32B is a bilingual large language model developed by LG AI Research, designed to serve as a high-performance hybrid system for enterprise and professional applications. Released as part of the fourth-generation EXAONE series, the model integrates two distinct operational modes: a Non-reasoning mode for rapid, knowledge-based responses and a Reasoning mode for complex multi-step problem solving. It supports Korean, English, and Spanish, with expanded capabilities for agentic tool use and function calling.
The model features a 32-billion parameter architecture that utilizes Hybrid Attention, combining local sliding-window attention with global attention in a 3:1 ratio. This design allows for efficient processing of an extended 128K token context window while maintaining computational performance. Additionally, the model implements a QK-Reorder-Norm scheme, which applies RMS normalization directly to the query and key projections to enhance stability and performance on downstream tasks compared to standard Transformer architectures.
Trained on 14 trillion tokens of high-quality data—including specialized technical, medical, and legal corpora—EXAONE 4.0 32B demonstrates a high level of domain expertise. In its Non-reasoning mode, it is optimized for high-speed instruction following, summarization, and translation. LG AI Research reports that the model has demonstrated proficiency levels sufficient to pass multiple professional certification exams, including the national medical doctor and customs broker examinations in South Korea.
EXAONE 4.0 32B is provided as an open-weight model for research and academic purposes. Its development followed a three-stage training process comprising large-scale supervised fine-tuning (SFT), reasoning-focused reinforcement learning (RL), and preference learning to ensure alignment with user instructions across various professional and creative domains.