EXAONE Deep is a series of reasoning-enhanced language models launched by LG AI Research, designed to improve reasoning capabilities in fields such as mathematics, science, and programming.
Key Features
- Multiple Model Sizes
The EXAONE Deep series offers three different model sizes:
- EXAONE Deep 32B: With 32 billion parameters, it excels at handling complex reasoning tasks.
- EXAONE Deep 7.8B: A lightweight version, delivering 95% of the 32B model’s performance with only 24% of its size.
- EXAONE Deep 2.4B: Optimized for mobile devices, achieving 86% of the 32B model's performance at just 7.5% of its size.
- Superior Reasoning Performance
EXAONE Deep demonstrates outstanding results across multiple benchmarks, especially in mathematics:
- MATH-500 benchmark: The 32B model scored 95.7.
- 2025 South Korean College Scholastic Ability Test (CSAT) (Mathematics section): Scored 94.5.
- 2024 American Invitational Mathematics Examination (AIME): Scored 90.0.
The 7.8B and 2.4B models also secured first place across all major benchmarks in science and programming.
- Open Source and Accessibility
LG AI Research released EXAONE Deep as an open-source project, allowing researchers and developers free access to explore, modify, and build on the model — driving AI innovation and adoption globally.
- Advanced Training Techniques
EXAONE Deep utilizes state-of-the-art training methods, including:
- Supervised Fine-Tuning (SFT)
- Direct Preference Optimization (DPO)
These techniques enhance reasoning performance, enabling the model to excel in complex tasks.
- Adaptability and Flexibility
EXAONE Deep is designed to be highly adaptable, supporting various frameworks and platforms such as TensorRT-LLM, vLLM, and Ollama — ensuring seamless deployment across diverse environments.
Application Scenarios
- Education and Learning
- Supports students in solving mathematical and scientific problems.
- Offers step-by-step solutions and logical reasoning breakdowns, helping learners grasp complex concepts more effectively.
- Programming and Software Development
- Generates code, debugs programs, and provides coding suggestions.
- Its reasoning capability allows it to understand programming logic and produce high-quality code snippets, making it ideal for software development and tech support.
- Chatbots and Customer Support
- Powers intelligent chatbots to deliver accurate, contextual responses to user queries — enhancing customer experience.
- Scientific Research and Data Analysis
- Handles complex data analysis tasks, helping researchers extract valuable insights from large datasets.
- Supports hypothesis validation and experimental design through its advanced reasoning abilities.
- Content Generation and Writing
- Assists with article writing, report creation, and creative content generation.
- Generates relevant, structured content based on user-provided topics — ideal for marketing and media industries.
- Multilingual Processing
- Supports multilingual applications, excelling in translation and understanding — especially between English and Korean.
- Provides value in international business and cross-cultural communication scenarios.
- Complex Problem Solving
- Excels in handling complex logical problems and decision-making processes.
- Suitable for finance, law, and medicine — fields that require high-level reasoning and precise analysis.
With cutting-edge performance, scalability, and open-source availability, EXAONE Deep positions itself as a powerful reasoning AI model — reshaping innovation across education, programming, research, and business solutions.