EXAONE 4.0 32B (Reasoning) is a hybrid large language model developed by LG AI Research, designed to integrate general-purpose linguistic proficiency with specialized, high-level reasoning capabilities. Released in July 2025, it serves as the flagship "professional" model in the EXAONE 4.0 series, succeeding the EXAONE 3.5 and EXAONE Deep lineages. The model is specifically optimized for complex problem-solving in expert domains such as law, medicine, accounting, and STEM fields.
The model utilizes a Hybrid Attention architecture that combines local sliding window attention with global full attention in a 3:1 ratio, enabling a context window of 128K tokens. To enhance global context understanding, the architecture omits Rotary Positional Embeddings (RoPE) for global attention and introduces a QK-Reorder-Norm scheme. This architectural modification applies LayerNorm directly to attention and MLP outputs while adding RMS normalization after the Q and K projections to improve training stability and downstream performance.
A defining feature of the model is its dual-mode operation, offering both Non-reasoning and Reasoning modes. The Reasoning mode utilizes a step-by-step thought process, initiated via a specialized <think> tag, which allows the model to form and verify hypotheses before generating a final answer. This approach is intended to reduce hallucinations and improve accuracy on multi-step tasks such as advanced mathematics and coding.
EXAONE 4.0 32B was pretrained on approximately 14 trillion tokens, a significant increase over its predecessors, to broaden its world knowledge and cross-lingual capabilities. In addition to its primary support for Korean and English, the model's multilingual scope was expanded to include Spanish. It also supports the Model Context Protocol (MCP) and native function calling, providing a foundation for integration into agentic workflows and tool-based environments.