SEOUL, South Korea – LG Electronics, a leading global technology company, has announced the launch of Korea’s first open-source AI model, EXAONE 3.0. This groundbreaking development signifies a major step forward for Korea in the global AI landscape, which has been dominated by U.S. tech giants, as well as emerging enterprises from China and the Middle East.
EXAONE 3.0: A Milestone for Korean AI
On August 7th, LG AI Research, a division of LG Electronics, revealed the EXAONE 3.0 model. This open-source AI model is based on the Decoder-only Transformer architecture and boasts a parameter count of 7.8 billion. The training data volume, measured in tokens, is a massive 8 trillion, making it a bilingual model designed for both English and Korean.
The announcement of EXAONE 3.0 highlights Korea’s growing presence in the global AI sector, which has traditionally been led by U.S. tech giants like Google and Microsoft, as well as Chinese and Middle Eastern companies like Huawei and ZTE.
EXAONE 3.0: A Model of Excellence
LG News reported that the EXAONE 3.0 model is equipped with a 7.8 billion instruction-tuning model that is being released for open-source use. The company aims to facilitate more meaningful research in the AI field and further enhance the AI ecosystem.
According to official tests, the English capabilities of the EXAONE 3.0 model have reached global top-level, with an average score in real-world examples that surpasses other leading models such as Llama 3.0 8B and Gemma 2 9B. In terms of mathematics and coding, EXAONE 3.0 also achieved top rankings, demonstrating strong reasoning abilities. Furthermore, the model achieved top rankings in Korean language tests, both in real-world examples and single benchmarks.
Performance Improvements and Cost Savings
LG claims that the EXAONE 3.0 model has improved its performance in various aspects compared to its predecessor. The model’s reasoning time has been reduced by 56%, memory usage has decreased by 35%, and operational costs have been lowered by 72%. Additionally, the cost of the model has decreased by 6% compared to the initial EXAONE 1.0 release.
Training and Expansion Plans
The EXAONE 3.0 model has been trained on 60 million professional data cases related to patents, code, mathematics, and chemistry. LG plans to expand the model’s training to encompass 100 million cases across various fields by the end of the year.
Conclusion
The launch of EXAONE 3.0 marks a significant milestone for Korea in the global AI domain. With its impressive performance and open-source nature, the model is poised to contribute to the advancement of AI research and development worldwide.
As Korea continues to assert its presence in the AI landscape, the success of EXAONE 3.0 could be a testament to the nation’s commitment to innovation and technological advancement.
Views: 0