Bllossom 8B

Bllossom 8B is a Korean-English bilingual language model developed by the Bllossom research team, with the Llama 3 based 8B checkpoint released in 2025 on Hugging Face under the MLP-KTLim organization. It continues the Bllossom line of research that started with Llama 2 based models in a published arXiv study on language augmentation for low-resource languages.

The model applies three techniques: Korean vocabulary expansion for better expressiveness, bilingual pretraining to link knowledge between Korean and English, and instruction tuning on custom Korean language-and-culture data with human feedback via DPO. The original Bllossom paper reported qualitative wins over other Korean models of the same size, with bilingual SFT showing a 93 percent preference win rate against Korean-only instruction tuning.

Weights are open and the underlying base carries Llama 3 community license terms, so commercial use requires checking the Llama 3 license conditions. The family also includes a 70B variant and a vision-language variant for multimodal Korean tasks.