LG Unveils AI 'Exaone 3.5' After 4 Months: "Understands 100 Pages of A4 at Once"

'ExaOne 3.5' 3 Types of Open Source Released
Top Performance Among Global Open Source AI Models
Director Baek Gyeonghun "AI, Speed Battle is Crucial"

Actual usage of 'Chat ExaOne' created based on LG's super-large AI 'ExaOne'. Photo by LG

On the 9th, LG AI Research Institute released version 3.5 of its generative AI model, ‘EXAONE’, as open source. This comes just four months after the domestic first release of EXAONE 3.0 as open source in August. Previously, LG first announced EXAONE 1.0 in December 2021 and then released EXAONE 2.0 in July last year, showing an accelerated pace in AI technology innovation by shortening the intervals between version releases.

Unlike the previous version which only released a general-purpose lightweight model, LG AI Research Institute this time unveiled three types of models: ultra-lightweight for on-device use, general-purpose, and high-performance. This allows various users including enterprises, institutions, academia, and startups to utilize the EXAONE models, lowering the barriers to technology use across the AI ecosystem and is expected to promote broader innovation and collaboration.

On November 21, Baek Gyeong-hoon, head of LG AI Research, presented the research achievements of the year at 'LG AI Insight 2024' held at LG Science Park in Magok, Gangseo-gu, Seoul. Provided by LG

The on-device model uses 2.4 billion parameters (numerical representations of rules created by AI while learning data) and is designed to run AI on lightweight devices such as mobile phones. The general-purpose model has 7.8 billion parameters and can be used flexibly according to user needs. Although the size is the same as the previous open-source version, its performance has improved. The high-performance model, based on 32 billion parameters, supports professional data analysis and large-scale tasks.

LG cited enhanced long-context understanding and processing capabilities as the greatest strengths of EXAONE 3.5. All these models can process 32,000 words at once, enabling them to handle long texts equivalent to 100 pages of A4 paper in a single pass. They demonstrated top-tier long-context understanding not only in English but also in Korean. EXAONE 3.5 applies Retrieval-Augmented Generation (RAG) and Multi-Step Reasoning (MSR) technologies to minimize hallucinations and significantly improve accuracy and reliability. Retrieval-Augmented Generation refers to a technology that improves answer accuracy by searching external data in real time.

LG AI Research Institute also released a technical report detailing the model’s performance evaluation results on the same day. EXAONE 3.5 recorded world-class performance in comparative evaluations against global open-source AI models from the US, China, and other countries. The ultra-lightweight on-device model ranked first on average across nine single benchmarks including math problem-solving and programming skills, solving problems more accurately and quickly than Meta’s ‘Llama 3.1’ and Google’s ‘Gemma 2’ of the same size.

Three Exaone 3.5 Models Released as Open Source. Provided by LG

Additionally, from that day, LG started the official service of ‘ChatEXAONE’, based on EXAONE, for its employees. ChatEXAONE enhances work efficiency through various functions such as real-time web search, document summarization, report writing, data analysis, and coding support. It guarantees the safety of internal data by applying information encryption and personal data protection technologies. By adding deep analysis and source selection features, it provides customized answers for 14 job roles and 133 specific tasks. LG plans to expand it to 128,000 tokens by the first half of next year. LG AI Research Institute will continuously improve ChatEXAONE’s functions based on employee feedback and aims to develop a self-acting AI agent by 2025.

LG AI Research Institute is targeting the development of AI agents based on Large Action Models (LAM) next year. Baek Gyeong-hoon, head of LG AI Research Institute, said, “With the rapid advancement of generative AI models recently, the speed of upgrades is crucial. We will accelerate innovation aiming for superintelligence applicable to industrial sites.”

LG AI Research Institute has formed global partnerships with 12 companies and institutions including Google Cloud, Dell Technologies, University of Michigan, Amazon Web Services, and NVIDIA. Domestically, it is discussing the application of EXAONE 3.5-based AI solutions to enterprise services with companies that have their own software such as Polaris Office and Hancom.

This content was produced with the assistance of AI translation services.