A 0.8-billion-parameter model translates 19 African languages offline

QVAC releases TranslatePsy-AfriSLM, a family of small open-weight models designed to translate between English and 19 Sub-Saharan African languages on computers and mobile devices.

Translating Hausa, Wolof, or Swahili without sending the text to a remote service. That is the promise behind TranslatePsy-AfriSLM, a family of models compact enough to run locally, including in regions where connectivity remains expensive or unreliable.

Developed by Tether AI Research and released under its QVAC initiative, the project is not limited to a single model. It includes three versions based on Qwen3.5, with 0.8 billion, 2 billion, and 4 billion parameters. Their weights, inference code, and data-processing scripts are publicly available.

The models cover English and 19 Sub-Saharan African languages: Afrikaans, Amharic, Hausa, Igbo, Kinyarwanda, Lingala, Luganda, Malagasy, Nyanja, Oromo, Shona, Somali, Southern Sotho, Swahili, Tswana, Wolof, Xhosa, Yoruba, and Zulu.

Translation primarily works in both directions between English and each of these languages. TranslatePsy-AfriSLM also retains conversational abilities: it can identify languages, respond in another language, and handle several translation requests within the same exchange instead of treating every sentence as an isolated operation.

The family is available in full precision and as lighter Q4 and Q8 GGUF versions intended for local use with llama.cpp. The 0.8B Q4 release is the most suitable configuration for devices with limited memory. It can run from a terminal or operate as a local server compatible with the OpenAI API format.

QVAC says this version can run on a laptop or smartphone without a specialized graphics card or continuous internet connection. The repositories do provide files intended for local execution, but the release does not include a ready-to-install mobile app, a list of supported phones, energy-consumption measurements, or speed comparisons across devices.

The technical ability to load the model on a phone therefore does not guarantee smooth performance on every device. Results will depend on available memory, the operating system, the selected version, and the software used to generate translations.

TranslatePsy-AfriSLM starts from a familiar observation: increasing model size does not necessarily compensate for a shortage of good training data. Languages with limited representation online have fewer parallel texts, meaning sentences that exist both in their original language and in translation. Available resources may also contain duplicates, errors, mixed languages, or incorrectly aligned sentences.

The team began by gathering 427 million raw sentence pairs from WMT22, MALA, OPUS, and Fine Translations. This collection covers the 19 target languages but is not being distributed with the rest of the project. The authors state that it was used for research purposes only.

A second source relies on synthetic data. Monolingual texts from MADLAD-400 were translated with NLLB-3.3B to produce new English-to-African-language pairs. The sentences were then cleaned, deduplicated, and checked against the evaluation sets to reduce the risk of test material appearing in the training data.

The main contribution lies in how these translations are selected. Instead of retaining everything available, the researchers combine three translation-quality estimators: AfriCOMET, SSA-COMET, and MetricX-24.

Their results are normalized using approximately 352,000 human-translated sentence pairs and combined into a single score. Pairs considered less reliable are removed, while the amount of data assigned to each language is capped to prevent better-resourced languages from dominating the collection.

The direction of the assessment also matters. A translation intended to train English-to-Swahili conversion must be evaluated in that same direction. Reversing the scoring direction produced an average decline of 12% under MetricX and 3.1% under SSA-COMET in the published experiments.

For the open datasets studied, this selection process reduced the training volume from 44.93 billion to 1.76 billion tokens, a 96% decrease, while preserving a comparable SSA-COMET score: 0.530 instead of 0.528 for the unfiltered collection.

That reduction does not describe the final volume used to train TranslatePsy-AfriSLM. The strongest configuration mainly relies on filtered synthetic material and contains 32.37 billion tokens. The 96% figure applies to an experiment conducted on the raw open-data collection, not to an equivalent reduction of the final training mixture.

The researchers also tested a combination of open and synthetic material. It reached 46.49 billion tokens but performed slightly worse than the 32.37-billion-token synthetic collection. Under their evaluation conditions, adding more weakly selected data eventually diluted the useful training signal.

Two additional collections are included during training. The first contains 4.6 million instruction examples, approximately half of them in African languages. Its purpose is to preserve conversational behavior, language identification, and the ability to follow requests that do not use a rigid translation command.

The second contains 24 million examples covering 38 European and Asian languages. Without this addition, specialized fine-tuning substantially damages some of the multilingual knowledge inherited from Qwen3.5. With it, the average decline outside the target African languages falls from 39% to 4.1% across the reported measurements.

The three models were fine-tuned for one epoch, with sequences capped at 2,048 tokens. This is also the context length used to evaluate the GGUF versions. TranslatePsy-AfriSLM is therefore aimed primarily at sentences, conversations, and short or segmented documents rather than translating an entire lengthy book in a single request.

The results cover Flores-200, BOUQuET, and Smol. These three datasets contain 1,012, 854, and 863 sentences, respectively. Performance is calculated across the 19 languages included in training, in both directions with English, and then averaged.

Under SSA-COMET, the 0.8B version scores 0.5944 on Flores-200, 0.6223 on BOUQuET, and