EXAONE-3.0-7.8B-Instruct

EXAONE-3.0-7.8B-Instruct is a bilingual (English and Korean) instruction-tuned language model developed by LG AI Research. Built on a 7.8-billion-parameter decoder-only Transformer with Grouped Query Attention (GQA), RoPE positional embeddings, and SwiGLU activations, it was pre-trained on 8 trillion curated tokens and post-trained with supervised fine-tuning and direct preference optimization (DPO).

The model targets real-world conversational and agentic use cases, with strong performance on instruction-following benchmarks such as MT-Bench, Arena-Hard, and WildBench. It particularly excels in Korean language tasks while remaining competitive with similarly sized open models (Llama 3.1 8B, Gemma 2 9B, Qwen2 7B) on English math, coding, and reasoning benchmarks including GSM8K, HumanEval, and MMLU-Pro.

Released in August 2024 under the EXAONE AI Model License Agreement 1.1 (non-commercial), EXAONE 3.0 is LG's first openly released model in the EXAONE family. It supports a 4,096-token context window and is available on Hugging Face for research and non-commercial applications.