Tether Releases Offline Translation Models for Edge Devices

What Tether released

Tether AI Research has announced new open-source translation models designed to run directly on smartphones, laptops and other edge devices.

The release includes TranslatePsy-AfriSLM models supporting 19 African languages, an AfriNano family supporting eight African languages and EuroNano models covering nine European languages. Tether says translations can run without an internet connection, keeping the processed input on the device instead of sending it to a third-party cloud service.

The model families are available through Hugging Face in full-precision and quantized configurations.

The access case

Local translation is particularly relevant where connectivity is unreliable, cloud inference is expensive or supported languages are poorly represented in mainstream products.

Tether positions the African-language models as infrastructure for education, agriculture, humanitarian communication and access to health information. Those use cases make device-level operation meaningful, but they also raise the standard for accuracy: a convenient translation is not necessarily a safe professional or clinical interpretation.

The benchmark boundary

Tether says its smallest 800-million-parameter AfriSLM model outperformed substantially larger systems across FLORES-200, BOUQuET and SMOL translation benchmarks. The company also says its highest-quality EuroNano model retained 98.4% of Meta NLLB-200 translation quality into English while using much less storage.

Those are company-reported results. Independent reproduction should examine performance language by language, across dialects and real-world inputs, rather than treating an aggregate benchmark as universal quality.

Privacy is a system property

Running a model locally removes the default need to transmit text to a cloud inference provider. That is a meaningful privacy advantage, especially for sensitive communications.

It is not a complete privacy guarantee. The surrounding application, telemetry, device security, model downloader and operating system can still expose data. The durable question is whether the full deployment remains local and auditable—not merely whether the model technically supports offline inference.

Sources and disclosures

Primary source: Tether, “Tether Releases Open-Source AI Translation Models for African and European Languages,” published September 2, 2026. Supported-language, storage and benchmark claims are attributed to Tether. Blue Rogues has not independently reproduced the benchmarks.

Related analysis

• [Pocket Bitcoin Breach Links Identities to Wallets](/news/pocket-bitcoin-breach-links-identities-to-wallets)
• [Polygon Discloses Patched Validator Risks](/news/polygon-discloses-patched-validator-risks)