Tether AI Research has introduced a new collection of open-source translation models designed to operate directly on smartphones, laptops and other everyday devices. The technology focuses heavily on underserved African-language communities, while a parallel European-language release aims to deliver compact, private translation without requiring continuous internet access.
The new QVAC TranslatePsy family includes TranslatePsy-AfriSLM, supporting 19 African languages, TranslatePsy-AfriNano, covering eight African languages, and TranslatePsy-EuroNano, which supports nine European languages.
Unlike cloud-based translation systems, these models are designed to process language locally. That means users can translate content without an internet connection, while their data remains on the device instead of being transmitted to external servers.
Bringing AI Translation to Underserved Communities
Language remains a major obstacle to accessing modern artificial intelligence across much of Africa. Many advanced AI systems primarily support widely spoken global languages and depend on reliable cloud connectivity, creating barriers for communities with limited internet access.
TranslatePsy-AfriSLM addresses this gap by supporting:
- Hausa
- Amharic
- Yoruba
- Lingala
- Swahili
- Igbo
- Zulu
- Somali
- Oromo
- Malagasy
- Kinyarwanda
- Xhosa
- Afrikaans
- Wolof
- Luganda
- Nyanja
- Shona
- Tswana
- Southern Sotho
Together, these languages cover communities across West, East, Central and Southern Africa, representing a substantial portion of the continent’s population.
Local translation could make educational courses, scientific resources, training materials and AI-powered learning applications available in languages that users already understand.
Smaller Models Deliver Competitive Translation Performance
One of the notable aspects of the release is the emphasis on model efficiency.
Tether says its 800-million-parameter TranslatePsy-AfriSLM model surpassed substantially larger systems in the FLORES-200, BOUQuET and SMOL translation evaluations.
The company attributes part of this performance to a quality-filtering technique designed to eliminate as much as 96% of poor-quality open-source training data.
Instead of simply increasing model size, the approach focuses on improving the quality of the information used during training.
If the results translate effectively into real-world deployments, smaller models could make advanced translation considerably easier to run on devices with limited processing power, memory and storage.
Healthcare Could Be a Major Use Case
Healthcare is one area where offline, local-language AI could have significant practical value.
People living in areas with unreliable connectivity may have difficulty accessing medical information, while healthcare workers often serve communities speaking multiple languages.
Tether says TranslatePsy-AfriSLM could potentially work alongside QVAC MedPsy, its smaller AI model focused on medical and healthcare applications, to make health information available in local languages.
Such technology could assist with health education and information access, although it would need appropriate safeguards. Translation and educational tools should not be treated as substitutes for qualified medical diagnosis or clinical care.
Applications Extend to Agriculture and Humanitarian Work
The potential uses go beyond classrooms and healthcare facilities.
Farmers could use local-language translation to access agricultural instructions, technical resources and educational material.
Humanitarian organizations could also benefit in regions where internet connectivity is limited. Field teams working across communities with different languages could use offline translation rather than depending on separate translation infrastructure or constant cloud access.
For NGOs and other organizations operating in remote areas, the ability to run translation directly on local hardware could make communication faster and more resilient.
Tether also points to its network of solar-powered kiosks across Sub-Saharan Africa. These facilities provide services such as phone charging and access to digital financial tools in locations where electricity and conventional banking infrastructure may be limited.
The company envisions such physical infrastructure potentially becoming additional access points for locally delivered educational and informational content.
Compact Translation Comes to Europe
Tether is applying the same local-first approach to European languages through TranslatePsy-EuroNano.
Rather than requiring numerous individual bilingual models, the European system uses compact multilingual models that can handle 90 translation directions across nine languages, using English as an intermediary language.
The smallest version requires approximately 36 MB of storage, according to Tether, compared with roughly 633 MB for an equivalent Firefox offline translation configuration.
That represents a claimed storage reduction of approximately 94%.
Tether also says its higher-quality EuroNano model retained 98.4% of Meta’s NLLB-200 translation quality when translating into English, while requiring substantially less storage.
The objective is to make capable translation models practical on ordinary consumer devices rather than limiting them to large cloud infrastructure.
Privacy and Offline AI at the Center of the Strategy
The release reflects a broader push toward local AI, where models operate directly on users’ hardware.
Running AI locally can provide several advantages, including reduced dependence on internet connectivity and greater control over potentially sensitive information.
For translation involving educational, financial, medical or personal material, keeping information on the device may also reduce the need to send that content to third-party cloud providers.
However, local AI comes with its own limitations, including device processing constraints, model-update challenges and differences in performance between hardware configurations.
Tether Positions QVAC as an Open AI Initiative
Paolo Ardoino, CEO of Tether, said the company’s objective is to prevent language and connectivity limitations from determining who can benefit from AI.
The company’s QVAC initiative is built around the concept of local and decentralized AI, with the goal of allowing intelligence systems to operate across a wide range of devices instead of concentrating computational power exclusively in corporate data centers.
TranslatePsy fits directly into that strategy by making multilingual AI smaller, portable and capable of operating without continuous cloud access.
The African models are available in multiple sizes, including 0.8B, 2B and 4B parameters, with both full-precision and quantized versions.
The TranslatePsy-Nano collection also includes the European and African Nano models.
Research Moves Toward Academic Review
The research behind TranslatePsy-AfriSLM has also been accepted for presentation at the EMNLP 2026 conference, providing an opportunity for the work to undergo broader academic discussion.
The release ultimately reflects a larger shift in AI development: rather than measuring progress only through increasingly large models, researchers are also focusing on efficiency, multilingual coverage, privacy and offline accessibility.
If compact translation models can deliver strong performance on everyday hardware, they could make AI substantially more accessible in communities where expensive cloud infrastructure and reliable internet remain out of reach. For Africa in particular, supporting a much wider range of local languages could become an important step toward expanding access to education, healthcare information, agriculture and digital services.
