TL;DR
- Tether launched open-source AI translation models that work offline on everyday devices for 19 African and 9 European languages.
- The AfriSLM model, with just 800 million parameters, outperformed much larger systems such as Qwen3.5-122B and TranslateGemma-27B across three benchmarks.
- A new filtering method eliminates up to 96% of low-quality training data, enabling higher performance with more compact models.
Tether announced the release of new families of open-source artificial intelligence translation models, designed to run directly on smartphones, laptops and other devices used in everyday life without requiring an internet connection.
The launch includes three models under the QVAC family: TranslatePsy-AfriSLM, which supports 19 African languages; QVAC TranslatePsy-AfriNano, covering 8 languages; and QVAC TranslatePsy-EuroNano, oriented toward 9 European languages.
The models process translations locally, meaning that user data remains on the device and is not sent to external servers. This decentralized architecture allows them to operate in areas with poor or no connectivity, a common condition in the sub-Saharan African communities that Tether’s project primarily targets.
Tether ❤️ Africa
Where you are born should never limit your potential.
True progress begins with access to stable money, stable energy, stable communications, and now also stable intelligence.Today, we introduce QVAC TranslatePsy / Afri SLM, a lightweight AI translation model… pic.twitter.com/cSc9sqylsy
— Tether (@tether) September 2, 2026
TranslatePsy-AfriSLM covers languages such as Hausa, Amharic, Yoruba, Swahili, Igbo and Zulu, among others, which together represent approximately half of the African population. Despite having just 800 million parameters in its smallest version, the model outperformed systems of far greater scale, including Qwen3.5-122B-A10B, TranslateGemma-27B and NLLB-3.3B, across the benchmarks FLORES-200, BOUQuET and SMOL.
Tether Implements a Filter That Radically Changes AI Training
A central technical element of this tool is the introduction of a quality estimation filtering method that discards up to 96% of open-source training data that presents low quality. Tether AI Research thus achieved higher translation performance with significantly smaller models, which reduces hardware requirements and makes their use viable on entry-level devices.
The intended applications span education, health, agriculture and humanitarian response. Combined with QVAC MedPsy, the company’s medical foundation model, TranslatePsy-AfriSLM could channel health information in local languages to hundreds of millions of people without access to conventional health systems. The solar kiosks Tether operates in sub-Saharan Africa, where residents can charge phones and access digital financial services, will serve as natural distribution points for these tools.
Democratization, Efficiency and Precision
In Europe, TranslatePsy-EuroNano unifies 90 translation directions into two compact models per performance tier. Its lightest version requires just 36 MB of storage, compared to the 633 MB of an equivalent offline Firefox configuration — a 94% reduction. The highest-quality model in the European family retained 98.4% of the accuracy of Meta’s NLLB-200 system when translating into English.
Paolo Ardoino, chief executive officer of Tether, stated that “language should not determine who can benefit from artificial intelligence” and noted that four billion people were excluded first from the traditional financial system and then from the most advanced technological tools. The models are available on Hugging Face and the research behind AfriSLM was accepted for presentation at the EMNLP 2026 conference.
