Introduction
Tether AI Research has released a new family of open-source translation models designed to make artificial intelligence more accessible to African users whose languages, devices and connectivity needs remain underserved by mainstream technology.
Announced on 2 September 2026, QVAC TranslatePsy-AfriSLM supports 19 African languages, while the smaller TranslatePsy-AfriNano supports eight. The models can run locally on compatible smartphones, laptops and other devices, allowing users to translate content without a continuous internet connection or sending their information to third-party cloud servers.
The launch shifts the conversation around African artificial intelligence from simply developing larger systems to building smaller, specialised models that can operate within the continent’s practical infrastructure constraints.
Nineteen African languages supported
TranslatePsy-AfriSLM supports Afrikaans, Amharic, Hausa, Igbo, Kinyarwanda, Lingala, Luganda, Malagasy, Nyanja, Oromo, Shona, Somali, Southern Sotho, Swahili, Tswana, Wolof, Xhosa, Yoruba and Zulu. Spanning West, East, Central and Southern Africa, these languages are spoken by communities that Tether estimates represent approximately half of the continent’s population. This coverage positions the model to support local-language education, healthcare information, agriculture, humanitarian response and cross-border communication across diverse African communities.
TranslatePsy-AfriSLM is available in three principal model sizes: 0.8 billion, 2 billion and 4 billion parameters. Full-precision and smaller quantised versions have also been published, allowing developers to select a model based on the storage, memory and processing capacity of their devices.
Smaller models compete with much larger systems
The most significant finding behind the launch concerns model efficiency.
According to the accompanying research paper, the smallest TranslatePsy-AfriSLM model contains approximately 800 million parameters but outperformed substantially larger systems, including TranslateGemma-27B and Qwen3.5-122B-A10B, on selected African-language translation tests.
The evaluation used benchmarks including FLORES-200, BOUQuET and SMOL, which assess translation performance across multilingual and lower-resource settings.
The comparison is significant because the Qwen model contains up to 122 billion total parameters, compared with AfriSLM’s 800 million. Parameter counts are not a complete measure of performance or operating cost, particularly because some models activate only part of their architecture for each task. Nevertheless, the difference demonstrates the potential value of models designed and trained specifically for African languages.
The results do not mean AfriSLM will outperform every larger model across all languages and real-world situations. They show that, on the benchmarks used in the study, a smaller specialised model can compete with and surpass much larger general-purpose systems.
The underlying research paper has been accepted for presentation at the 2026 Conference on Empirical Methods in Natural Language Processing.
Data quality rather than data volume
A central innovation in the project is its approach to training-data quality.
African-language datasets collected from open online sources can contain inaccurate translations, duplicated information, inconsistent spelling and unrelated sentence pairs. Training models on this material can increase computing requirements without necessarily improving translation quality.
Tether AI Research developed a quality-estimation filtering method that removed up to 96% of the available training tokens without reducing model performance. The remaining information was combined with African-focused synthetic data and curated human-quality translations.
This suggests that developing effective models for underrepresented languages may depend more on improving the relevance and accuracy of training data than continually increasing model and dataset size.
The researchers found that filtered synthetic data produced a stronger balance between translation quality and computational efficiency. This approach could help reduce the hardware, memory and energy requirements associated with developing and operating African-language translation systems.
Why offline processing matters in Africa
Most leading AI services depend on cloud infrastructure. Users send information through the internet, the request is processed in a remote data centre and the result is returned to their device.
This structure creates challenges where connectivity is unreliable, mobile data is expensive or users are handling sensitive information.
TranslatePsy processes translations directly on a user’s device. This means translation can continue without an active internet connection, personal information does not have to leave the device and smaller models can potentially operate on less powerful hardware.
These capabilities could make translation technology more practical for schools, health-information programmes, agricultural projects, humanitarian organisations and field operations serving communities outside major urban centres.
Education, healthcare and agricultural applications
Education represents one of the clearest potential applications. Schools, publishers and technology providers could use the models to translate educational materials, scientific information and digital learning resources into languages that learners understand more comfortably.
Healthcare organisations could potentially combine TranslatePsy-AfriSLM with specialised systems such as Tether’s QVAC MedPsy to provide general health education in local languages.
However, machine translation in healthcare requires significant safeguards. Errors involving symptoms, medication, dosage or treatment could cause harm. The models should therefore support supervised health education and communication rather than replace medical professionals or qualified interpreters.
In agriculture, the technology could help translate information about weather, crop management, pests, soil health and market access. Humanitarian organisations could also use offline translation to communicate in areas where internet connectivity has been disrupted or is unavailable.
These remain prospective applications. The published research demonstrates technical performance, but it does not yet provide evidence from large-scale deployments in African schools, clinics or humanitarian operations.
European models demonstrate the storage advantage
Tether AI Research has also released TranslatePsy-EuroNano, covering nine European languages and supporting 90 translation directions through English as a pivot language.
The smallest European deployment reportedly requires 36MB of storage, compared with 633MB for an equivalent Firefox offline-translation configuration. This represents an approximate 94% reduction in storage requirements.
According to the published model documentation, the system retains up to 98.4% of Meta’s NLLB-200 translation quality when translating into English. Its single-checkpoint deployment is also reported to be 56.7 times smaller and to require 3.53 times less peak memory during CPU deployment.
Although these European results should not automatically be applied to the African models, they demonstrate the broader engineering objective: making translation available through compact systems instead of depending entirely on large cloud-based infrastructure.
Open-source access could support local development
The models and supporting resources are available through Hugging Face. The TranslatePsy-AfriSLM collection includes the 0.8B, 2B and 4B models, together with quantised alternatives and supporting datasets.
Open access gives African developers, universities and organisations an opportunity to test the models, examine their limitations and adapt them for specific industries and communities.
However, open-source availability alone will not guarantee meaningful adoption. Effective deployment will require local testing, representative speakers, dialect coverage, terminology validation and continuous feedback from the communities the technology is intended to serve.
A step towards more accessible African AI
TranslatePsy-AfriSLM provides evidence that African-language artificial intelligence does not always require the largest possible model. Through language specialisation, aggressive data filtering and compact deployment options, the research demonstrates that smaller systems can deliver competitive translation performance.
Its coverage of 19 African languages, 800-million-parameter entry model and ability to remove up to 96% of noisy training tokens make the release technically significant.
The larger test will be whether these benchmark results translate into reliable performance across Africa’s diverse dialects, institutions and operating environments. If that validation happens, offline translation could expand access to education, information and digital services while giving users greater control over their data.
For now, the release represents a meaningful contribution to the development of smaller, open and locally deployable AI systems for African languages.