Alibaba Unveils Babel: A Multilingual Language Model on Hugging Face

Alibaba has released Babel, a new multilingual large language model (LLM) on Hugging Face, designed to serve over 90% of global speakers. The announcement was made via a social media post by Alibaba's representative, highlighting the introduction of two model variants: Babel-9B and Babel-83B.

Babel-9B is optimized for efficient single-GPU inference and fine-tuning, while Babel-83B sets a new standard for open multilingual LLMs. The models have been extensively evaluated on multilingual tasks, demonstrating superior performance compared to other open LLMs of similar size. Babel-9B-Chat leads among 10B-sized LLMs, and Babel-83B-Chat performs comparably to GPT-4o on certain tasks.

The models employ a unique layer extension technique to expand capacity without compromising performance, focusing on languages that are often underrepresented in existing models. This approach allows Babel to improve accuracy in languages with limited training data, achieving up to 10% better performance in these languages compared to previous multilingual LLMs.