LG AI Research Releases EXAONE 3.5 Open-Source Models

Following the release of EXAONE 3.0 last August, LG AI Research reflected feedback requesting models suited to various purposes and has released 3 open-source models based on EXAONE 3.5.

This lineup consists of three models depending on use case.

• 2.4B model: An ultra-lightweight model for on-device and low-spec GPU environments
• 7.8B model: A lightweight model for general-purpose use (improved performance compared to the previous version)
• 32B model: A frontier-AI-level high-performance model for customers who prioritize high performance

EXAONE 3.5 went through a Pre-training stage that removes duplicates and personal information from the dataset to improve model training efficiency, and strengthened instruction-following ability and reflection of user preferences through SFT (Supervised Fine-tuning) and DPO (Direct Preference Optimization).

In particular, it has excellent Long Context Understanding ability, supporting an effective context length of up to 32K tokens. This is highly advantageous when using RAG (Retrieval-Augmented Generation) technology, and it has demonstrated top-tier performance compared to global models in both English and Korean.