LG unveils multimodal AI model that outperforms global rivals

LG AI Research's EXAONE 4.5 / Courtesy of LG AI Research

LG AI Research introduced EXAONE 4.5, a multimodal artificial intelligence (AI) model designed to understand and reason across both text and images, marking a significant step in the company’s push to expand its proprietary AI ecosystem.

The new model integrates a self-developed vision encoder with a large language model into a single system, forming a vision-language model capable of processing complex documents such as contracts, technical drawings and financial statements with a high degree of accuracy.

According to LG AI Research, EXAONE 4.5 outperformed competing models, including those from OpenAI and Alibaba, across a range of benchmarks measuring visual understanding and reasoning. The model recorded an average score of 77.3 across five STEM-related evaluations, surpassing GPT-5 mini, Claude Sonnet 4.5 and Qwen3 235B.

Across 13 evaluation metrics — including general visual understanding and document-based reasoning — the model also exceeded the performance of GPT-5 mini, Claude Sonnet 4.5 and Qwen3-VL, the company said.

In coding performance, EXAONE 4.5 scored 81.4 on the LiveCodeBench v6 benchmark, outperforming Google’s Gemma 4, which scored 80.0. The model also demonstrated strong capabilities in analyzing complex charts, achieving 62.2 on the ChartQA Pro benchmark.

An LG AI Research official said the high benchmark scores indicate that the model goes beyond simple recognition of text and unstructured data, demonstrating an ability to understand context and respond to questions.

The model contains 33 billion parameters, roughly one-seventh the size of the previously released K-EXAONE, yet achieves comparable performance in text understanding and reasoning. The gains were made possible through a hybrid attention structure and multi-token prediction-based inference technology developed by the company.

LG AI Research said the model supports multiple languages, including Korean, English, Spanish, German, Japanese and Vietnamese.

The release is part of a broader effort to expand LG’s AI foundation model project, known as K-EXAONE, with the long-term goal of developing “physical intelligence” systems capable of understanding and interacting with the real world.

The institute also made EXAONE 4.5 available on Hugging Face for research, academic and educational use, continuing its push to foster a more open AI ecosystem.

Lee Jin-sik, head of EXAONE Lab at LG AI Research, said the model signals the company’s transition into a multimodal AI era. “Starting with this model, we will expand AI’s range of understanding to include voice, video and physical environments, creating systems that can make practical decisions and take action in industrial settings,” he said.

Kim Myoung-shin, head of the Trust and Safety Office at LG AI Research, said the company is working to develop AI that deeply understands Korea’s history, culture and social context. “While more AI systems are becoming proficient in Korean, understanding historical and cultural sensitivities is a different level of challenge,” he said.

This article was published with the assistance of generative AI and edited by The Korea Times.