Tether releases open-source AI translation models for African and European languages

Tether’s AI research unit just shipped something that has nothing to do with stablecoins and everything to do with the fact that most of the world’s languages are essentially invisible to modern AI. The company released an open-source neural machine translation model family called TranslatePsy-AfriSLM, covering 19 Sub-Saharan African languages with a dataset of roughly 215 million bidirectional training examples paired with English.

The models are tiny. We’re talking 21 to 35 MB per language pair, small enough to run directly on a phone without ever pinging a server. That’s the point: translation that works offline, keeps data private, and doesn’t require the kind of cloud infrastructure that rarely exists in the regions where these languages are actually spoken.

What Tether actually built

The project, developed by Tether’s QVAC division, includes both the dataset and the translation models themselves. The dataset, called TranslatePsy-AfriSLM-Synthetic-Mix, is hosted on Hugging Face under a CC BY-NC 4.0 license, meaning researchers and developers can freely use it for non-commercial purposes.

Languages covered include Swahili, Yoruba, and Amharic, among 19 Sub-Saharan African languages total.

The models are built to be Bergamot-compatible, which matters because Bergamot is the open-source translation engine that Mozilla uses in Firefox’s built-in translation feature. Compatibility with that ecosystem means these models can slot into existing infrastructure without developers needing to rebuild from scratch.

Speed is the other headline number. The models process a sentence in approximately 46 milliseconds, which QVAC claims makes them up to 78 times faster than larger models like Salamandra-2B.

The QVAC SDK and European language coverage

Alongside the African language models, Tether released the QVAC SDK, a software development kit that bundles various AI models into a unified deployment framework. European language support comes through this SDK rather than through separate dedicated datasets, giving developers a single integration point for multilingual translation.

According to QVAC’s documentation, the SDK can extend coverage to dozens or even hundreds of languages using approximately 50 language pairs to cover 26 languages.

The accompanying research paper has been submitted to EMNLP 2026 and is available on arXiv (arXiv:2608.18655). The code and models are documented on GitHub under Tether’s AI research organization at tether-ai-research/qvac-translatepsy-afri-slm.

Why a stablecoin company is building translation AI

Tether’s stablecoin sees massive adoption across Africa, Southeast Asia, and Latin America, regions where local currencies are volatile, banking infrastructure is thin, and where many of the world’s low-resource languages are spoken. Building AI tools that work in those languages, on devices common in those markets, and without requiring persistent internet connections is a play that reinforces Tether’s presence in exactly the communities where its stablecoin already has traction.

By releasing everything under permissive licenses and submitting the research for academic review, Tether is positioning QVAC as a legitimate AI research outfit. Whether that perception sticks will depend on whether the models actually perform well in real-world deployments, something the NLP research community will be testing now that the weights are public.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.