KORMo

KORMo-10B is a 10.8-billion-parameter Korean-English bilingual model from KAIST's MLP and NLPCL labs with SeoulTech, published in October 2025. It is the first large-scale attempt at a fully open Korean LLM trained predominantly on synthetic data: 68.74 percent of the Korean training corpus was carefully curated synthetic text.

The paper's key finding is that synthetic data, when balanced across linguistic coverage and instruction styles, can sustain long-horizon pretraining without model collapse or degradation. The resulting model matches contemporary open-weight multilingual baselines on reasoning, knowledge, and instruction-following benchmarks, with near-native Korean reasoning after bilingual instruction tuning.

Everything is public: checkpoints, datasets, training recipes, and logs are released at huggingface.co/kormo-lm. It is as much a reproducibility artifact as a product, setting a template for building transparent models in low-resource language settings.