LG Launches EXAONE 3.5: Open-Source AI Model & Enterprise Service

Seoul, South Korea - LG AI Research released an open-source version of its EXAONE 3.5 language model on Monday, accompanied by the launch of an enterprise AI agent service for LG employees.

The new release comes four months after EXAONE 3.0. 

It includes three variants: an ultra-lightweight model for on-device use (2.4 billion parameters), a lightweight general-purpose model (7.8 billion parameters), and a high-performance specialized model (32 billion parameters).

The company claims top performance among open-source AI models across 20 benchmarks, particularly in real-world usability, long text processing, coding, and mathematics.

EXAONE 3.5, the latest release, boasts unique features such as enhanced Retrieval-Augmented Generation technology and Multi-step Reasoning capabilities. 

These are specifically designed to reduce hallucination and increase answer accuracy, setting it apart from its predecessors and other models in the market. 

The model can process approximately 100 pages of text in a single instance.

ChatEXAONE is an enterprise service that LG says is equipped with encryption and privacy protection technologies for secure use within the company.

In addition, the company added in the press release that ChatEXAONE is not just a standard AI agent.

LG evaluated it as a useful tool for enterprises because it includes complex query analysis and user-customized search functions. 

LG AI Research has announced plans to expand the model's token capacity from 32,000 to 128,000 in the first half of 2025, demonstrating our dedication to continuous improvement.

The research division is already developing a Large Action Model (LAM) for release in 2025, a project that promises to revolutionize the field of AI.