Artificial intelligence is usually associated with massive data centers, cloud computing and increasingly large models. Tether is taking a different approach: making specialized AI small enough to run directly on everyday devices.

Tether AI Research has released a new family of open-source translation models designed to run locally on smartphones, laptops and other edge devices. The project places particular emphasis on African languages that have historically received limited support from mainstream AI systems, while a separate compact model family targets nine European languages.

The most important part of the announcement, however, may not be the number of languages. It is the fact that the models are designed to work offline, without requiring every translation request to be sent to a cloud server.

Tether's New TranslatePsy Models Explained

The new translation technology is part of Tether's broader QVAC artificial intelligence initiative. The September 2, 2026 release includes several TranslatePsy model families with different priorities around language coverage, model size and device requirements.

Model	Coverage	Main Focus
TranslatePsy-AfriSLM	19 African languages	Higher-quality translation on local devices
TranslatePsy-AfriNano	8 African languages	Compact, browser-scale deployment
TranslatePsy-EuroNano	9 European languages	Very small footprint for local translation

Tether says TranslatePsy-AfriSLM is available in 0.8B, 2B and 4B parameter versions, with full-precision and quantized builds. The company says the models can run through llama.cpp on smartphones and laptops without an expensive GPU, internet connection or account.

The 19 languages supported by AfriSLM are Afrikaans, Amharic, Hausa, Igbo, Kinyarwanda, Lingala, Luganda, Malagasy, Nyanja, Oromo, Shona, Somali, Southern Sotho, Swahili, Tswana, Wolof, Xhosa, Yoruba and Zulu.

Tether says these languages span West, East, Central and Southern Africa and collectively represent roughly half of the continent's population.

Why Offline AI Translation Matters

Most popular AI applications depend on cloud infrastructure. A user's request is transmitted to a remote server, processed by a large model and then returned over the internet.

That model works well when connectivity is fast and affordable. It becomes much less practical when internet access is unreliable, expensive or unavailable.

Tether's TranslatePsy approach is different. Translation can take place directly on the user's device, meaning the system can continue operating without a live internet connection. Tether also says processing locally keeps users' data on their devices rather than sending it to third-party cloud servers.

That creates two potentially important advantages: availability and privacy.

A student, healthcare worker, researcher or field worker may not need a permanent connection to use the translation technology. At the same time, local processing can reduce the amount of sensitive text that has to leave the device.

Of course, offline AI is not automatically better in every situation. Device hardware, storage, model quality, language coverage and battery consumption remain important considerations.

Africa Has a Major AI Language Gap

The language problem is particularly significant in Africa because the continent has extraordinary linguistic diversity. UNESCO estimates that Africa is home to around 2,000 languages, yet only a small fraction have a meaningful presence in digital services such as localized software, translation and speech technologies.

UNESCO has also highlighted the difficulties that African-language users can face with AI systems trained primarily on dominant languages. In other words, simply making an AI model available globally does not necessarily mean that it works equally well for every language community.

That is the gap Tether is attempting to address with TranslatePsy-AfriSLM.

Instead of trying to build one enormous general-purpose model, the project focuses on specialized translation models for languages that have historically had less AI training data and fewer digital resources.

The 800 Million Parameter Model Is the Interesting Part

One of the most notable claims from Tether concerns the smallest AfriSLM model.

TranslatePsy-AfriSLM-0.8B contains 800 million parameters.

Tether says that despite its relatively small size, the model outperformed much larger systems including Qwen3.5-122B-A10B, TranslateGemma-27B and NLLB-3.3B across the FLORES-200, BOUQuET and SMOL translation benchmarks.

Tether's own research says the smallest model can outperform systems up to 152 times its parameter count on the evaluated benchmarks.

These are Tether-reported benchmark results, rather than independent verification. That distinction matters, particularly when comparing models with very different architectures, training methods and evaluation setups.

Still, the underlying idea is important: specialized AI does not necessarily need to be enormous to be useful.

Data Quality May Matter More Than Model Size

Tether says one of the biggest challenges in developing African-language translation models was not simply computing power. It was the quality of available training data.

According to Tether, its quality-estimation pipeline evaluates candidate sentence pairs using multiple quality-estimation models and keeps the strongest training signals.

The company reports that the filtering process reduced its open-source training data from approximately 44.93 billion tokens to 1.76 billion tokens—a reduction of about 96%—without a meaningful loss in translation quality.

Tether argues that adding enormous amounts of low-quality data can eventually dilute the useful training signal rather than improve the model.

That is an important lesson for low-resource languages. When high-quality human-translated datasets are limited, better filtering and carefully generated synthetic data can potentially compensate for the lack of massive datasets.

Translation Speed: About 46 Milliseconds per Sentence

Model size is only useful if the model can actually respond quickly enough on the hardware people own.

Tether's QVAC research reports that a Bergamot-based translation setup can process a sentence in approximately 46 milliseconds on a Linux laptop in its benchmark.

The same research tested a compact translation model on a Pixel 10 Pro XL. In batch mode, Tether reported approximately 26.1 milliseconds per sentence and around 870 tokens per second.

Tether says that at that speed, a full page of text can be translated in under three seconds in batch mode, entirely on-device and without network calls.

These numbers are benchmark measurements, not guarantees for every smartphone or laptop. Real-world performance will vary according to hardware, language pair, model configuration and workload.

Test	Reported Result
Bergamot on Linux laptop	~46 ms per sentence
Pixel 10 Pro XL batch mode	26.1 ms per sentence
Pixel 10 Pro XL throughput	~870 tokens/second

Bergamot Compatibility Could Matter for Firefox

There is another technical detail that deserves attention: Bergamot.

QVAC's translation stack supports Bergamot as an inference engine. Bergamot is associated with the offline translation technology used in Mozilla's Firefox translation ecosystem.

That compatibility is important because it means QVAC's local translation architecture is not completely isolated from an existing browser-based offline translation ecosystem.

QVAC's technical documentation lists both its own local inference stack and Bergamot as translation options. Developers can load supported neural machine translation models and perform translation locally through the QVAC SDK.

This does not mean that Tether has announced a new official Firefox integration for every TranslatePsy model. It means that the underlying QVAC translation stack supports the Bergamot inference technology, which creates an interesting path for developers already working with that ecosystem.

QVAC SDK Launched Earlier in 2026

The translation release is easier to understand when viewed alongside Tether's earlier QVAC work.

In April 2026, Tether launched the QVAC SDK, an open-source, cross-platform software development kit intended to provide a common interface for local AI capabilities.

The SDK brings together technologies including llama.cpp, whisper.cpp, Bergamot, Stable Diffusion and ONNX Runtime through a unified architecture.

QVAC's documentation describes the system as a way to build local and peer-to-peer AI applications, allowing models to perform inference directly on devices rather than relying exclusively on cloud infrastructure.

TranslatePsy is therefore not an isolated experiment. It is another example of the local-first AI strategy Tether has been developing through QVAC.

Healthcare Could Become a Major Use Case

Language barriers can become particularly serious when the information being communicated involves healthcare.

Tether says TranslatePsy-AfriSLM could work alongside QVAC MedPsy, its smaller medical-focused AI model, to make health information more accessible in local languages.

The concept is straightforward: a medical information system could potentially operate locally on a device and use translation technology to communicate information in a user's preferred language.

However, translation AI should not be treated as a replacement for qualified medical professionals. Errors in medical terminology or context can have serious consequences.

The strongest potential use case is therefore likely to involve AI-assisted communication with appropriate human oversight rather than fully autonomous medical decision-making.

Education, Agriculture and Humanitarian Work

Healthcare is only one possible application.

Education could benefit from local-language translation as well. A large amount of digital educational material is produced in a relatively small number of dominant languages. Translation into more African languages could make some of that material easier to access.

Agriculture is another potential area. Information about weather, crops, pests, farming methods and safety procedures is useful only if people can understand it.

Humanitarian organizations could also potentially use offline translation in areas where internet infrastructure is damaged or unavailable.

The common factor is connectivity. An AI model that can operate locally does not have to stop working simply because the internet connection disappears.

What About the Nine European Languages?

Tether's release also includes TranslatePsy-EuroNano, a compact translation family designed for nine European languages.

Tether says TranslatePsy-Nano is designed for a browser-scale footprint, with the Nano family covering nine European and eight African languages.

The company has emphasized reducing the amount of storage required to deploy translation models locally.

Tether's research reports that one European-language deployment can require roughly 36 MB of storage, compared with approximately 633 MB for an equivalent Firefox offline translation configuration in the company's comparison.

That represents roughly a 94% reduction in storage requirements, according to Tether's benchmark.

Again, these figures are based on Tether's own testing and configuration. Actual storage requirements can vary depending on the model, language pair and software environment.

Why Smaller AI Models Are Becoming More Important

The AI industry has spent years competing to build larger models with more parameters and greater computing requirements.

But another trend is developing at the same time: AI is moving closer to the device.

Smartphones, laptops and other consumer hardware are becoming capable of running increasingly sophisticated models locally.

Local AI can offer several advantages:

Less dependence on an internet connection
Potentially better privacy for sensitive information
Lower cloud infrastructure requirements
Faster responses when network latency is significant
More opportunities for AI use in remote locations
Greater control for developers and users

The trade-off is that local models generally have fewer resources than the largest cloud systems. Developers therefore have to balance model size, accuracy, speed, storage and battery consumption.

Tether's Approach Is Different From Simply Building a Bigger Model

TranslatePsy illustrates a broader change in how AI performance can be measured.

Instead of asking only how large a model is, developers can ask whether the model solves a particular problem efficiently on hardware that people already own.

Translation is a good example because it is a specialized task. A translation model does not necessarily need the broad reasoning abilities of a general-purpose AI assistant.

If a smaller specialized model can produce high-quality translations while using dramatically less memory and computing power, it can become easier to deploy at scale.

That is particularly relevant for low-resource languages, where the primary challenge may not be building the world's largest model but creating a model that understands the language well enough to be genuinely useful.

Open Source Could Help the Project Expand

Another important part of the release is its open-source nature.

Tether says the TranslatePsy-AfriSLM models, inference code and related data-processing resources are being made available so researchers and developers can experiment with the technology.

The AfriSLM models are available in three sizes—0.8B, 2B and 4B parameters—with quantized versions available for smaller deployments.

This gives developers more flexibility. A powerful laptop may be able to use a larger model, while a smartphone or low-resource device could use a smaller quantized version.

The open-source approach also leaves room for developers to build applications Tether itself may not have anticipated, including education tools, local-language assistants and offline communication software.

What This Means for the Future of AI

Tether's translation project highlights a part of the AI industry that is sometimes overshadowed by the race for ever-larger models.

The next stage of AI may not be defined only by bigger data centers and more powerful chips. It may also be defined by smaller models that can run locally and solve specific problems efficiently.

For Africa, that could have particular importance because language support remains one of the biggest barriers to digital inclusion.

UNESCO has highlighted how many African languages remain poorly represented online. Tether's approach does not solve that problem by itself, but it offers another technical path: train specialized models for those languages and make them small enough to run on devices people already own.

The Bottom Line

Tether's latest AI release is more significant than a simple multilingual translation announcement.

It combines open-source models, offline processing, small parameter counts, local inference and support for historically underserved languages.

TranslatePsy-AfriSLM supports 19 African languages, while the broader TranslatePsy-Nano family covers nine European and eight African languages.

Tether's reported benchmarks suggest that relatively small models can compete with much larger systems in specialized translation tasks. The company also reports major reductions in training-data volume and storage requirements, as well as fast local inference on laptop and smartphone hardware.

But the most important test will happen outside benchmark tables.

If students, healthcare workers, farmers, researchers and field teams can reliably use these models on ordinary devices without an internet connection, the technology could become a practical tool for digital inclusion.

The broader lesson is simple: the future of AI may not only be about making models bigger. It may also be about making them smaller, faster, more private and available in more people's own languages.

Frequently Asked Questions

What did Tether AI Research release?

Tether AI Research released open-source TranslatePsy translation models designed to run locally on smartphones, laptops and other edge devices.

How many African languages does TranslatePsy-AfriSLM support?

TranslatePsy-AfriSLM supports 19 African languages, including Hausa, Swahili, Yoruba, Igbo, Zulu, Somali, Oromo and others.

Can Tether's translation models work offline?

Yes. Tether says the TranslatePsy models are designed to process translations locally without requiring an internet connection.

What is TranslatePsy-Nano?

TranslatePsy-Nano is Tether's compact translation family covering nine European and eight African languages, with an emphasis on a browser-scale footprint and local deployment.

What is Bergamot?

Bergamot is an inference technology used in offline neural machine translation. QVAC's translation documentation supports Bergamot as one of its translation inference engines, connecting the QVAC stack with technology used in the Firefox offline translation ecosystem.

How fast is Tether's offline translation technology?

Tether's published benchmark reports approximately 46 milliseconds per sentence for a Bergamot setup on a Linux laptop. A Pixel 10 Pro XL test reported 26.1 milliseconds per sentence in batch mode. Actual performance will vary by device and configuration.

What is the QVAC SDK?

The QVAC SDK is Tether's open-source, cross-platform development framework for local and peer-to-peer AI applications. It was launched in April 2026 and supports multiple AI capabilities, including translation.

Why is offline AI important?

Offline AI can continue working where internet connectivity is unreliable. It can also reduce the need to send sensitive information to remote cloud servers.

Are Tether's benchmark results independently verified?

No independent verification is implied by Tether's announcement. The benchmark figures discussed in this article are identified as company-reported results. Independent testing and real-world deployments will provide additional evidence about the models' performance.

Is Tether's AI project connected to USDT?

The translation project comes from Tether AI Research and its QVAC initiative. It is part of Tether's broader technology strategy and is separate from the core operation of the USDT stablecoin.

Sources

Editorial Note: This article is based primarily on Tether's official announcement, QVAC technical documentation and published research material. Performance figures and benchmark comparisons attributed to Tether should be understood as company-reported results unless independently verified.

Disclaimer: This article is provided for informational purposes only. It does not constitute financial, medical, investment or technical advice.

Tether Brings Offline AI Translation to 19 African and 9 European Languages