LG AI EXAONE Demonstrates Multimodal Expansion Capabilities

LG AI Research unveiled 'EXAONE 4.5' on the 9th, a multimodal AI model capable of understanding both images and text. This showcases the expansion of its self-developed AI model, 'K-EXAONE,' to process visual information beyond text. 'EXAONE 4.5' is a Vision-Language Model (VLM) that integrates its self-developed vision encoder and a Large Language Model (LLM) into a unified structure. It has been defined as a preparatory stage for evolving EXAONE into 'Physical Intelligence,' capable of understanding and interacting with the physical world beyond virtual environments.

This technological development is part of its participation in the Ministry of Science and ICT's 'Domestic AI Foundation Model (Dokpamo)' project, aligning with the government's policy direction to reduce reliance on foreign AI and secure independent technological prowess. Through the Dokpamo project, LG AI Research developed the LLM 'K-EXAONE' with 236 billion parameters and received the highest score in the first stage evaluation. Following the second evaluation in August, if advancement to the third stage is confirmed, plans are in place to expand 'K-EXAONE' into a multimodal AI that understands images, speech, and video.

Despite having 33 billion parameters, 'EXAONE 4.5' achieves equivalent performance to 'K-EXAONE' in text comprehension and reasoning. By applying its self-developed hybrid attention structure and high-speed inference technology, it has demonstrated superior benchmark results compared to global competing models. 'EXAONE 4.5' shows strengths in inferring industrial documents such as contracts and technical drawings. It supports multiple languages, including Spanish and German, beyond Korean and English.

LG AI Research is simultaneously pursuing accessibility expansion efforts for the realization of 'AI for All,' including public release on Hugging Face, specialized learning for Korean culture, and utilization of a youth AI expert nurturing program. The release of open weights and the operation of educational programs by LG AI Research are expected to bridge the gap between model release and actual utilization, contributing to the popularization of such technologies.

Kim Myung-shin, Head of the Trust and Safety Office at LG AI Research, stated, "While AI with Korean language capabilities is increasing, understanding historical and cultural sensitivities is a different level of challenge." He added, "EXAONE will evolve into an AI that simultaneously secures rich expressiveness and reliability, based on its self-designed AI risk classification system (K-AUT)."