Tether Data, S.A. de C.V., operating as Tether AI Research, published TranslatePsy-AfriSLM on August 27, 2026. The release presents TranslatePsy-AfriSLM as an open-source machine-translation suite for 19 African languages, with model weights, inference code and processing scripts licensed under Apache 2.0 for research and education.
The synthetic training data follows a different licence: CC-BY-NC 4.0, also limited to research and education. That distinction matters for anyone considering the models for a product or service. The announcement provides access to the core materials for the stated non-commercial purposes, but it does not grant unrestricted commercial permission.
A local route for 19 languages
TranslatePsy-AfriSLM covers Afrikaans, Amharic, Hausa, Igbo, Kinyarwanda, Lingala, Luganda, Malagasy, Nyanja, Oromo, Shona, Somali, Southern Sotho, Swahili, Tswana, Wolof, Xhosa, Yoruba and Zulu.
The release also announces GGUF versions for llama.cpp. This route is intended to support inference on a laptop or smartphone without a costly GPU, continuous internet access or advanced hardware. It gives researchers and educators a way to work with the models closer to where the translation is needed, although the deployment option does not alter the research-and-education limits attached to the published materials.
From hundreds of millions of sentence pairs to a filtered training set
The open-source data pool contains 427 million raw sentence pairs gathered from WMT22, MALA, OPUS and Fine Translations. According to the release, this mixture is reserved for research and is not intended for publication.
The team created a separate synthetic mixture from monolingual text in MADLAD-400. It used NLLB-3.3B through CTranslate2 to generate parallel data in both directions: English to African languages and African languages to English.
Quality scoring combines AfriCOMET, SSA-COMET and MetricX-24. The unified score was calibrated against about 352,000 human-translated sentence pairs. After filtering, the open-source training budget fell from 44.93 billion tokens to 1.76 billion, a stated 96% reduction without a reported drop in translation performance.
The final synthetic configuration used a unified quality-estimation threshold of z̄ ≥ 0.68. It contained 32.37 billion tokens and reached 0.632 SSA-COMET on BOUQuET with the 2B model. These figures describe the data-selection process and the reported evaluation setup; they do not establish one universal configuration for every language or deployment.
What the reported evaluations show
The evaluation covered 19 languages seen during fine-tuning and eight held-out languages. The described setup used Flores-200, BOUQuET and Smol, with 1,012, 854 and 863 sentences respectively.
For the 19 languages seen during fine-tuning, the release reports that TranslatePsy-AfriSLM-0.8B outperformed Qwen3.5-122B-A10B and TranslateGemma-27B across all three benchmarks. It matched NLLB-3.3B on Flores-200 and exceeded it on BOUQuET and Smol.
The held-out-language results are more specific. TranslatePsy-AfriSLM-2B improved SSA-COMET over the Qwen3.5-2B backbone in every one of the eight languages, with particularly strong gains for Sepedi, Bambara and Akan. The improvement was less consistent on the other metrics, so the strongest conclusion is limited to the named metric and evaluation setup rather than a universal ranking of translation systems.
TranslatePsy-AfriSLM therefore combines broad coverage, published model artifacts and a path toward local inference. The practical compromise is clear: the announced access is aimed at research and education, while the performance claims remain tied to specific benchmarks, metrics and model sizes. Anyone assessing the release for broader deployment must account for both limits.