LG AI Research has announced the release of EXAONE 3.0, a 7.8 billion parameter instruction-tuned language model, marking a significant advancement in the field of large language models (LLMs). This is the first open model in the EXAONE family, aimed at democratising access to expert-level artificial intelligence capabilities.
The release is intended to foster innovation and collaboration within the AI community by providing a high-performance model for non-commercial research purposes.
EXAONE 3.0 has been extensively evaluated across a variety of benchmarks, demonstrating competitive real-world performance and instruction-following capabilities. It excels particularly in Korean, while also achieving strong results in English across general tasks and complex reasoning.
The model’s architecture is based on a decoder-only transformer with advanced features like rotary position embeddings and grouped query attention, supporting a maximum context length of 4,096 tokens.
The model was trained using a diverse dataset, ensuring robust performance in real-world scenarios. It features a bilingual tokeniser designed to optimise performance in both English and Korean, addressing the linguistic complexities of the Korean language.
The training process involved extensive pre-training and post-training techniques, including supervised fine-tuning and direct preference optimisation, to enhance the model’s instruction-following capabilities.
The EXAONE 3.0 model is available on Hugging Face for research purposes, supporting the broader AI community’s efforts in developing innovative applications. LG AI Research aims to integrate advanced AI into everyday life, making expert knowledge accessible to a wider audience. The model is expected to contribute significantly to advancements in Expert AI, particularly in bilingual environments.
LG AI Research has emphasised the importance of responsible AI development, conducting thorough compliance reviews and ethical assessments to minimise risks associated with data usage. The model has been evaluated for potential social and ethical issues, with measures in place to ensure its safe and ethical deployment.
