LG EXAONE launched: Korea's first open source AI model

LG AI Research has announced the launch of a completely new open source AI model called EXAONE 3.0. This is Korea's first open source AI model, also an achievement marking the country's official entry into the global AI field, which is currently dominated by giant technology companies from Korea. America, along with a series of emerging names in China. EXAONE 3.0 currently only supports two language options - English and Korean - but will be continuously expanded in the future.

EXAONE 3.0 is an open source model, developed based on the Transformer architecture using only a decoder. This model boasts up to 7.8 billion parameters and over 8T of training data (tokens). Talking about the launch of EXAONE 3.0, LG's press release said:

Among the EXAONE 3.0 language models built for various purposes, the 7.8 billion parameter command tuning model is being pre-sourced so that it can be used for research purposes. We hope the release of this model will help AI researchers both domestically and abroad conduct more meaningful research and help the AI ​​ecosystem move one step further on the path to development and dissemination. update.

LG's internal tests show that EXAONE 3.0's English capabilities have reached "the highest level in the industry". The average score of real-world use cases is said to be at the top, surpassing some famous current models such as Llama 3.0.

When it comes to mathematical computation and coding, EXAONE 3.0 also tops the average score and has strong reasoning capabilities. EXAONE 3.0 ranked first in average scores for both real-world use cases and single benchmarks in Korea.

Notably, EXAONE 3.0 has demonstrated superior performance and cost efficiency compared to its predecessors, EXAONE 1.0 (December 2021), and EXAONE 2.0 (July 2023). This model is said to reduce inference time by 56%, memory usage by 35% and operating costs by 72% compared to the previous generation model.

To date, EXAONE 3.0 has been trained on 60 million specialized data instances related to patents, codes, mathematics and chemistry. The company plans to expand the training data to 100 million cases in various fields by the end of this year.

To reduce power consumption when operating the model, LG AI Research will focus on researching optimization technologies and making the above EXAONE 3.0 more compact and lightweight. The company has succeeded in reducing model size by up to 97% while increasing performance compared to EXAONE 1.0.