LG AI Research has introduced South Korea’s first open-source AI model, EXAONE 3.0. This model marks South Korea’s entry into the competitive global AI landscape, currently led by American and emerging companies from China and the Middle East. With 7.8 billion parameters and trained on 8 trillion tokens, this model aims to advance AI research and development, providing a new tool for researchers both locally and internationally. Below we shall talk more about South Korea’s first Open Source AI model in further detail.
Table of Contents
South Korea’s first Open Source AI model:
EXAONE 3.0 is an open-source AI model built by LG AI Research. It’s designed with the Decoder-only Transformer architecture and comes packed with 7.8 billion parameters, making it a powerful tool in the AI field. The model was trained on 8 trillion tokens, which gives it a strong foundation for understanding and generating text in both English and Korean.
LG decided to open-source the 7.8B model ahead of its full lineup to help researchers. They believe this will encourage meaningful AI research and push the AI ecosystem forward, both in South Korea and globally.
Impressive Performance:
In tests conducted by LG, EXAONE 3.0 has shown impressive results, particularly in its English language abilities. It has been ranked as one of the top models in the world, even outperforming other well-known models like Llama 3.0. When it comes to tasks like mathematical calculations and coding, EXAONE 3.0 also ranks first, demonstrating strong reasoning capabilities.
In Korea, the model has excelled, ranking first in both practical use cases and single benchmarks. It’s also more efficient than its predecessor, reducing inference time by 56%, memory usage by 35%, and operating costs by 72%.
Extensive Data training:
The model has been trained on 60 million cases of professional data, including patents, codes, mathematics, and chemistry. LG plans to expand this training data to 100 million cases by the end of the year, making the model even more versatile.
LG AI Research has also focused on making the model more energy-efficient. They succeeded in reducing the model’s size by 97% compared to EXAONE 1.0, while still improving performance. This makes EXAONE 3.0 a more sustainable and powerful tool for AI research and applications.
