Babel-9B

Babel is an open multilingual large language model from DAMO Academy researchers that targets the top 25 languages by number of speakers, covering more than 90 percent of the global population including many languages neglected by other open models such as Bengali, Urdu, Swahili, and Javanese. Instead of plain continued pretraining, the team grew the model through a layer-extension technique to raise its performance ceiling.

Two variants were released: Babel-9B for efficient inference and fine-tuning, and Babel-83B as the flagship. The paper reports Babel-9B-Chat leading among 10B-sized models on multilingual tasks. Exact license terms should be confirmed on the model card before commercial deployment.