EXAONE 4.0 – LG’s Multimodal Reasoning Foundation Model

What is EXAONE 4.0?

EXAONE 4.0 is a hybrid reasoning foundation model independently developed by LG AI Research in South Korea. It integrates general natural language processing with advanced reasoning capabilities and supports Korean, English, and Spanish. The model comes in two versions: a 32B professional-grade version and a lightweight 1.2B on-device version. The 32B version is fine-tuned on several national certification exams and is designed for highly specialized domains. The 1.2B version is compact yet powerful, supports local deployment, and is ideal for privacy-sensitive and security-critical use cases. EXAONE 4.0 has demonstrated outstanding performance on challenging international benchmarks, scoring 81.8 on MMLU-Pro and 85.3 on AIME 2025, showcasing its strong ability to handle complex tasks.

Key Features of EXAONE 4.0

• Advanced Reasoning Capabilities:
  EXAONE 4.0 excels in handling complex tasks in science, mathematics, and programming. It supports step-by-step reasoning and logical thinking, enabling it to solve difficult problems.

• Multilingual Support:
  The model supports Korean, English, and Spanish, enhancing its applicability on a global scale.

• Function Calling and MCP Interface:
  EXAONE 4.0 supports function calling and the Model Context Protocol (MCP), providing foundational support for agent-based applications and making it easier to integrate with other systems.

• Professional and On-Device Versions:

  • Professional Version (32B): Fine-tuned on six national certification exams in fields such as law, accounting, and medicine. Ideal for high-expertise domains.

  • On-Device Version (1.2B): Compact and capable of local deployment. Suitable for scenarios requiring high privacy and security.

• Educational and Commercial Use:
  Available for free use by educational institutions. A commercial API is also offered for businesses to integrate and deploy quickly.

Technical Architecture of EXAONE 4.0

• Hybrid Reasoning Framework:
  EXAONE 4.0 combines general NLP with advanced reasoning abilities. It is designed to handle complex tasks using a step-by-step, logic-based reasoning approach.

• Deep Learning and Neural Networks:
  Built on deep learning, particularly the Transformer architecture, EXAONE 4.0 is trained on large-scale datasets for optimal performance. It is proficient at understanding and generating natural language in challenging contexts.

• MCP and Function Calling Support:
  With Model Context Protocol (MCP) and function calling, the model can interact with external systems and tools to automate more complex tasks.

• Optimization and Compression:
  The on-device version is optimized using model compression techniques, reducing its size by 50% while maintaining high performance. This makes it suitable for deployment on resource-constrained devices.

• Multilingual Training:
  Trained on multilingual datasets, the model supports Korean, English, and Spanish, making it broadly applicable across regions and industries.

Project Resources for EXAONE 4.0

• Official Website: https://www.lgresearch.ai/blog/view?seq=575

• Technical Paper: https://www.lgresearch.ai/data/cdn/upload/EXAONE_4_0.pdf

Application Scenarios for EXAONE 4.0

• Intelligent Customer Service:
  Rapidly responds to customer inquiries in multiple languages with accurate answers to complex questions, enhancing customer satisfaction.

• Educational Support:
  Generates practice questions, grades assignments, and offers personalized learning suggestions, supporting customized education.

• Healthcare and Medicine:
  Assists doctors in diagnosis, provides medical knowledge lookup, and helps patients better understand medical recommendations, improving healthcare efficiency.

• Programming Assistance:
  Can generate code snippets, debug code, and offer programming suggestions, significantly improving developer productivity.

• Enterprise Operations:
  Automates report generation, data organization, meeting scheduling, and more, streamlining workflows and boosting office efficiency.