Tether AI Research has released three families of open-source translation models designed to run directly on smartphones, laptops and other edge devices without requiring an internet connection or cloud processing.
QVAC TranslatePsy-AfriSLM supports 19 African languages, AfriNano covers eight, and EuroNano handles nine European languages. Local inference keeps text on the device, allowing translations to continue in areas with unreliable connectivity while avoiding the need to send user data to third-party cloud servers.
AfriSLM Covers Languages Used Across Half of Africa
TranslatePsy-AfriSLM supports Hausa, Amharic, Yoruba, Lingala, Swahili, Igbo, Zulu, Somali, Oromo, Malagasy, Kinyarwanda, Xhosa, Afrikaans, Wolof, Luganda, Nyanja, Shona, Tswana and Southern Sotho.
Those languages span West, East, Central and Southern Africa and are spoken across regions representing roughly half of the continent’s population.
Running the models locally could make translated educational material, agricultural information and health resources accessible on ordinary consumer hardware in areas where continuous cloud access is impractical. The same local-processing architecture has been a central part of QVAC’s development, including its BitNet LoRA framework for training and running AI models on consumer GPUs and smartphones.
800M-Parameter Model Beats Far Larger Rivals
The smallest TranslatePsy-AfriSLM model contains about 800 million parameters but exceeded Qwen3.5-122B-A10B, TranslateGemma-27B and NLLB-3.3B across the FLORES-200, BOUQuET and SMOL African translation benchmarks.
Tether’s training process filtered out as much as 96% of lower-quality open-source parallel data before model training. AfriSLM is available in 0.8-billion, 2-billion and 4-billion parameter versions, with full-precision and quantized releases for different hardware limits.
The models use the Apache 2.0 license, allowing developers to download, modify and integrate them into other applications. The underlying AfriSLM research has also been accepted for presentation at EMNLP 2026.
EuroNano Shrinks Offline Translation to 36MB
TranslatePsy-EuroNano supports German, Spanish, French, Italian, Portuguese, Finnish, Czech, Dutch and Swedish, using English as a pivot across 90 translation directions.
Its smallest deployment requires about 36MB of storage, roughly 94% less than the 633MB required by an equivalent Firefox offline translation setup. The highest-quality EuroNano configuration retained 98.4% of Meta’s NLLB-200 translation quality when translating into English.
The release expands a QVAC strategy that already includes local inference, model fine-tuning and developer funding for software designed to operate without centralized cloud providers. TranslatePsy-AfriSLM, AfriNano and EuroNano are available in full-precision and quantized versions for local deployment.
