Tether AI Research has launched the open-source TranslatePsy translation family through its QVAC initiative, covering 19 African languages and nine European languages across QVAC TranslatePsy-AfriSLM, QVAC TranslatePsy-AfriNano and QVAC TranslatePsy-EuroNano. The models are designed to run locally on smartphones, laptops and other edge devices, enabling translation without an internet connection while keeping user data on the device. The African release includes the TranslatePsy-AfriSLM-Synthetic-Mix dataset, containing roughly 215 million bidirectional training examples paired with English. Its models range from 21MB to 35MB per language pair, and QVAC says AfriSLM processes a sentence in approximately 46 milliseconds, up to 78 times faster than Salamandra-2B. AfriSLM contains 800 million parameters and, in Tether's internal evaluations, outperformed Qwen3.5-122B-A10B, TranslateGemma-27B and NLLB-3.3B in FLORES-200, BOUQuET and SMOL benchmarks; those results are the company's own verification and have not been confirmed by third parties. Tether AI Research said a quality-estimation filtering method removed up to 96% of low-quality open-source training data. TranslatePsy-EuroNano supports 90 translation directions, with its smallest deployment requiring 36MB of storage, about 94% less than an equivalent Firefox offline-translation configuration, while its highest-quality model retained 98.4% of Meta's NLLB-200 quality when translating into English. The release positions the tools for education, healthcare, agriculture, humanitarian response and cross-border communication in areas with limited connectivity. The African dataset is hosted on Hugging Face under a CC BY-NC 4.0 license, while the models and code are documented through Tether's AI research organization. The associated research has been accepted for presentation at EMNLP 2026, although earlier Tether News material and the existing record described it as submitted, creating a publication-status discrepancy.