TranslatePsy-AfriSLM brings on-device translation to 19 African languages

Hugging Face Blog has introduced TranslatePsy-AfriSLM, an open-source machine-translation suite designed for 19 Sub-Saharan African languages. Published August 27, 2026, the project focuses on making useful language tools available on everyday devices, including in regions with limited connectivity and costly cloud access.

The release covers Afrikaans, Amharic, Hausa, Igbo, Kinyarwanda, Lingala, Luganda, Malagasy, Nyanja, Oromo, Shona, Somali, Southern Sotho, Swahili, Tswana, Wolof, Xhosa, Yoruba and Zulu. Its first model release is tied to a forthcoming EMNLP 2026 paper on improving data quality for low-resource translation.

Smaller models, leaner training

The team says its smallest model, TranslatePsy-AfriSLM-0.8B, matches or beats much larger systems on three reported benchmarks: FLORES-200, BOUQuET and Smol. The comparison includes models such as Qwen3.5-122B-A10B and TranslateGemma-27B, with the article stating that the 0.8B system performs competitively against models up to 152 times larger.

A major part of the approach is filtering noisy and repetitive data. Quality-estimation methods reduced the open-source training set from 44.93 billion tokens to 1.76 billion—a 96% cut—without reported degradation in translation results. The project also uses synthetic parallel data and supports multi-turn translation, language identification and cross-lingual conversation.

For builders, the release includes model weights, inference code, processing scripts and GGUF versions for llama.cpp. That makes laptop and smartphone deployment more practical, potentially enabling offline education, agricultural information and multilingual communication. The weights and code use Apache 2.0, while synthetic training data is available under CC-BY-NC 4.0, limiting commercial use.

Source: Hugging Face Blog