EXAONE 4.0: Unified Large Language Models Integrating Non-reasoning and Reasoning Modes

2025-07-15

AI Summary

Existing large language models (LLMs) often sacrifice either interactive usability (e.g., low-latency response) or complex reasoning capability, and lack efficient multilingual support and cross-platform deployment (e.g., servers and mobile devices). Method: We propose a unified dual-mode LLM architecture that jointly optimizes non-reasoning (interaction efficiency) and reasoning (complex task handling) capabilities. Key innovations include a dynamic mode-switching mechanism, a lightweight tool-calling interface, multilingual (English, Korean, Spanish) joint pretraining, and coordinated model compression across scales. Contribution/Results: (1) The first open-source dual-mode unified architecture balancing high responsiveness and strong reasoning; (2) Two publicly released variants—32B (server-grade) and 1.2B (on-device)—achieving state-of-the-art performance on multilingual understanding, tool utilization, and reasoning benchmarks; (3) Full model weights and inference code released on Hugging Face, advancing practical agent-oriented AI deployment.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Computer Vision: Large Vision Models

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

This technical report introduces EXAONE 4.0, which integrates a Non-reasoning mode and a Reasoning mode to achieve both the excellent usability of EXAONE 3.5 and the advanced reasoning abilities of EXAONE Deep. To pave the way for the agentic AI era, EXAONE 4.0 incorporates essential features such as agentic tool use, and its multilingual capabilities are extended to support Spanish in addition to English and Korean. The EXAONE 4.0 model series consists of two sizes: a mid-size 32B model optimized for high performance, and a small-size 1.2B model designed for on-device applications. The EXAONE 4.0 demonstrates superior performance compared to open-weight models in its class and remains competitive even against frontier-class models. The models are publicly available for research purposes and can be easily downloaded via https://huggingface.co/LGAI-EXAONE.

Problem

• Integrates Non-reasoning and Reasoning modes for enhanced usability and reasoning
• Extends multilingual support to include Spanish alongside English and Korean
• Offers two model sizes for high performance and on-device applications

Innovation

• Unified model with reasoning and non-reasoning modes
• Supports multilingual including English, Korean, Spanish
• Offers 32B and 1.2B model sizes