New lightweight translation models bring machine translation efficiency to African languages

A team of researchers has introduced AfriNLLB, a collection of compressed machine translation models optimized for translating African languages efficiently. The work addresses a critical gap in language technology by making translation capabilities practical for deployment in resource-limited environments across the African continent.

Scope and languages supported#

AfriNLLB covers 15 language pairs spanning 30 translation directions, encompassing languages spoken across Africa and official languages of the African Union. The models handle translation between English and 13 African languages—including Swahili, Hausa, Yoruba, Amharic, Somali, Zulu, Lingala, Afrikaans, and Wolof—as well as bidirectional translation between French and Lingala and Wolof. The system additionally supports Arabic (Modern Standard Arabic), French, Portuguese, and Spanish to facilitate broader multilingual communication.

Technical approach#

The researchers built AfriNLLB by starting with the NLLB-200 600M base model and then reducing its size through iterative layer pruning and quantization. This compression strategy allows the models to run faster without dramatically sacrificing translation quality. To further refine performance on African language pairs, the team applied knowledge distillation from a larger teacher model and fine-tuned the compressed versions on parallel text corpora they assembled specifically for African languages.

Release and availability#

The project delivers two distinct versions of AfriNLLB: a Transformers version that researchers can continue to customize and fine-tune, and a CTranslate2 version optimized for fast inference during deployment. Notably, the team has also released all training data used in model development, enabling the broader research community to build upon this foundation and create improved translation systems.

Impact and evaluation#

Testing reveals that AfriNLLB achieves translation performance on par with the baseline NLLB model while delivering substantially faster processing speeds. This efficiency gain makes the technology viable for real-world applications in settings where computational resources are constrained—a critical consideration across many parts of Africa where deployment infrastructure may be limited.