In September 2021 researchers at Naver published “What Changes Can Large-scale Language Models Bring? Intensive Study on HyperCLOVA,” describing an 82-billion-parameter generative model trained on a Korean-centric corpus of 560 billion tokens. It was one of the first publicly documented large language models on the GPT-3 scale built outside the United States and trained primarily on a non-English language.
The paper showed that with Korean-specific tokenization and training, HyperCLOVA achieved state-of-the-art in-context zero-shot and few-shot results on Korean tasks, demonstrating that the GPT-3 recipe could be reproduced for another language by a national internet company rather than only by US labs. The work was presented at EMNLP 2021 and became the foundation for Naver’s later HyperCLOVA X model and its sovereign-AI strategy.
Why a business reader should care: HyperCLOVA was an early proof that countries could build frontier-scale language models in their own languages, an idea that has since hardened into national policy across Asia, the Middle East, and Europe.