Korea Gets Its First Open Source AI Model Thanks to LG

South Korea has debuted in the global AI arena with EXAONE 3.0, the country’s first open-source AI model. Developed by LG AI Research, this launch positions South Korea alongside established American tech giants and emerging players from China and the Middle East in the competitive field of artificial intelligence.

EXAONE 3.0 is built on the Decoder-only Transformer architecture and boasts impressive specifications, including 7.8 billion parameters and training on 8 terabytes of data. Currently, the model supports English and Korean languages.

LG said in a press release:

Among the EXAONE 3.0 language model lineup built for various purposes, the 7.8B instruction adjustment model is being open-sourced in advance so that it can be used for research.

By making this model open-source, LG AI Research aims to foster innovation and advance the AI ecosystem globally. The company envisions researchers in South Korea and internationally leveraging EXAONE 3.0 to conduct groundbreaking research and push the boundaries of AI technology.

LG AI Research reports that EXAONE 3.0’s English language capabilities have reached top-tier global standards. In company-conducted tests, the model’s average score in real-world applications reportedly outperformed several competitors, including Meta’s latest Llama 3.0 AI model.

The model’s prowess extends beyond language processing, with EXAONE 3.0 demonstrating exceptional performance in mathematical computations and coding tasks. Its strong reasoning abilities have contributed to its top-ranking average scores in these areas.

On the domestic front, EXAONE 3.0 has claimed the leading position in both practical use cases and individual benchmarks within South Korea. Compared to its predecessor, the new model boasts significant efficiency improvements, with a 56% reduction in inference time, 35% less memory usage, and a substantial 72% decrease in operational costs.

LG AI Research has significantly enhanced EXAONE 3.0’s knowledge base, training it on an impressive 60 million instances of specialized data encompassing patents, coding, mathematics, and chemistry. The company has ambitious plans to further broaden this dataset, aiming to incorporate 100 million cases across diverse fields by year’s end.

LG AI Research has prioritized energy efficiency in EXAONE 3.0’s operation. Through intensive research into optimization techniques and model streamlining, the team has achieved a staggering 97% reduction in model size compared to EXAONE 1.0. Notably, this dramatic downsizing has not come at the cost of performance; instead, the model’s capabilities have been enhanced.