LG AI Research has launched its highly anticipated EXAONE 3.5, a powerful open-source AI model that pushes the boundaries of AI performance and usability. This new release, coming just four months after EXAONE 3.0, introduces a suite of improvements including long text processing, coding capabilities, and groundbreaking privacy features. EXAONE 3.5 includes three models catering to different use cases, from ultra-lightweight on-device models to high-performance specialized solutions.
- Breakthrough Open-Source Models
- EXAONE 3.5 comes with three open-source models:
- Ultra-Lightweight Model (2.4 billion parameters) for on-device use
- Lightweight Model (7.8 billion parameters) for general-purpose tasks
- High-Performance Model (32 billion parameters) for specialized applications
- These models enable a wide range of use cases, offering flexibility from mobile devices to high-demand enterprise environments.
- EXAONE 3.5 comes with three open-source models:
- Enhanced Privacy and Performance with RAG and Multi-step Reasoning
- LG AI Research has advanced its Retrieval-Augmented Generation (RAG) technology, enabling real-time web search results and document uploads for more accurate, reliable responses.
- Multi-step Reasoning further boosts EXAONE 3.5’s accuracy by enabling AI to break down complex queries and generate logical, step-by-step answers.
- Unmatched Performance Across 20 Benchmarks
- EXAONE 3.5 leads globally across 20 benchmark tests, showcasing exceptional performance in real-world usability, long text processing, coding, and mathematical tasks.
- A technical report published by LG AI Research provides transparency, with individual and average scores for these benchmarks.
- ChatEXAONE: The AI Agent for LG Employees
- LG employees now have access to ChatEXAONE, an enterprise AI agent built on EXAONE 3.5.
- ChatEXAONE integrates information encryption and privacy protection, offering functionalities like real-time web search, document summarization, translation, and coding without compromising security.
- New features, Deep and Dive, allow for comprehensive multi-stage answers and the ability to select specific search scopes like general, global, academic, or YouTube for precise answers.
- Future Prospects: LAM and AI Evolution
- Looking ahead, LG AI Research is exploring Large Action Models (LAM), which will enable AI to plan and act autonomously.
- This development is set to culminate in the creation of AI agents capable of independent operation, with a projected launch in 2025.
EXAONE 3.5 marks a significant milestone in the evolution of AI, offering unparalleled flexibility, security, and performance for both research and enterprise applications. LG AI Research’s open-source approach accelerates innovation in the AI ecosystem, providing a foundation for future advancements like LAM technology. With a global network of collaborators and an expanding ecosystem, LG is committed to driving AI technology that enhances national competitiveness and real-world applications.
