LG AI Research has introduced “EXAONE 3.0,” South Korea’s inaugural open-source AI model, positioning the country among the leading nations in artificial intelligence development. This model, based on the Decoder-only Transformer architecture, features 7.8 billion parameters and is trained on 8 terabytes of data, currently supporting both English and Korean languages. LG aims to use EXAONE 3.0 to foster global innovation and contribute to the AI ecosystem.

 

In company tests, the model demonstrated superior English language skills, outperforming competitors like Meta’s Llama 3.0, with notable abilities in mathematical calculations and coding. The new version also shows enhanced efficiency, with a 56% reduction in inference time, 35% less memory usage, and a 72% cut in operational costs compared to its predecessor.

 

LG AI Research has enriched EXAONE 3.0’s knowledge base by training it on 60 million specialized data instances in various fields and plans to expand to 100 million instances by the end of the year. The model’s size has been reduced by 97% without compromising performance, emphasizing LG’s commitment to energy efficiency and innovation in AI technology.