EXAONE 3.0 - LG's Open-Source AI Model Designed for English and Korean

What is EXAONE 3.0

EXAONE 3.0 is an open-source AI model released by LG AI Research, with 780 million parameters, designed for English and Korean. EXAONE 3.0 performs excellently in both English and Korean language tests, ranking first in real-world use cases and mathematical coding. Compared to its predecessor, EXAONE 3.0 has significantly optimized inference speed, memory usage, and operating costs. It has currently been trained on 60 million professional data cases, with plans to expand to 100 million cases by the end of the year. The model can be accessed on the Hugging Face platform.

Features of EXAONE 3.0

• Bilingual Support: Designed for English and Korean, capable of handling natural language processing tasks in both languages.
• High Performance: Outstanding performance in multiple English and Korean tests, including real-world use cases and mathematical coding capabilities.
• Open Source: Model code and training data are public, facilitating use and further research by researchers and developers.
• Optimized Efficiency: Compared to the previous generation, inference time is reduced by 56%, memory usage by 35%, and operating costs by 72%.
• Professional Domain Training: Trained on 60 million data cases in professional domains such as patents, code, mathematics, and chemistry.

Technical Principles of EXAONE 3.0

• Decoder-only Transformer Architecture: This architecture is a variant of the Transformer model, containing only a decoder without an encoder. The model is more direct and faster when generating text because the decoder can directly generate the output sequence.
• Large-Scale Parameters: EXAONE 3.0 has 780 million parameters, enabling the model to capture more complex language patterns and relationships, thereby improving its ability to understand and generate text.
• Bilingual Model: Trained for English and Korean, capable of handling natural language understanding (NLU) and natural language generation (NLG) tasks in both languages.
• Massive Training Data: The model uses 8TB of training data, including a wide range of language materials, helping to improve the model's generalization ability and accuracy.

Project Address of EXAONE 3.0

Application Scenarios of EXAONE 3.0

• Language Translation: Supports translation tasks between English and Korean, helping cross-language communication and information transfer.
• Text Generation: Can be used to generate creative writing, news articles, social media content, etc.
• Question-Answering System: Build an intelligent question-answering system to provide users with accurate and quick information feedback.
• Text Summarization: Automatically generate summaries of documents or web pages, helping users quickly grasp the main content.