LG Group has announced the release of its new AI model, EXAONE 3.0, on August 7, marking a pivotal development in the AI sector. This latest version of LG’s large language model (LLM) offers improved performance and reduced size compared to its predecessor, EXAONE 2.0. In a groundbreaking move for a Korean company, EXAONE 3.0 has been made open-source, enabling widespread access and fostering innovation across the AI research community.
According to the LG AI Research Institute, the focus was on developing lightweight and optimized technologies, successfully reducing the model’s size to one-third of its original while enhancing its performance. This open-source release is aimed at bolstering AI research and strengthening national AI competitiveness.
EXAONE 3.0 has been trained on over 60 million specialized data entries, covering areas such as patents, code, mathematics, and chemistry, with plans to expand to over 100 million entries by year-end, including legal, biomedical, and medical sciences. The model boasts a 56% decrease in inference processing time, 35% less memory usage, and 72% lower operating costs compared to EXAONE 2.0. It achieved a score of 74.1 in Korean language evaluation, outperforming competitors like Alibaba’s Qwen-2 and Meta’s LLaMA 3.1, and exceeded Meta and Google’s AI models in several areas including mathematics and coding.
The LG AI Research Institute emphasized the significance of this release as the first self-developed AI model from Korea to be open-sourced, providing opportunities for research institutions and startups to leverage cutting-edge technology. The development of EXAONE began in 2020 with the establishment of the LG AI Research Institute, progressing through versions 1.0 and 2.0 before reaching this latest iteration.
