Uganda-based nonprofit Sunbird AI has released Sunflower v2, an open suite of AI models that it says expands language coverage from 31 Ugandan languages to 67 across Africa. The release adds speech recognition, text-to-speech and a lightweight model intended to run offline on mid-range smartphones, extending the project beyond text translation and generation.

Sunbird says the suite includes a larger language model for translation and comprehension, a smaller model designed to run on laptops, speech-to-text coverage for 51 African languages and text-to-speech for 20. The organisation says the models and supporting datasets are openly available for researchers and developers to use and adapt. It describes the release as a pipeline that can take spoken language into text and produce spoken output. (Sunbird AI release notes)

Open tools, local constraints

Speech technology can make digital services more accessible to people who are more comfortable speaking than typing, and local processing can matter where connectivity is unreliable. Open weights and data can also let African researchers test, adapt and scrutinise tools instead of relying only on closed services.

The practical details matter. Sunbird’s release describes one compact model as offline-capable on mid-range phones and demonstrates a related on-device learning tool in Luganda and Acholi. That is not the same as showing every feature works offline across all 67 languages or on every phone. The distinction between model availability, language coverage and a finished consumer product is important.

Benchmarks are a starting point

Sunbird reports strong performance on African-language benchmarks, but these comparisons come from the organisation behind the models. Independent evaluation across dialects, accents, noisy recordings and real user tasks will be needed to understand how well the tools work outside curated tests.

The v2 release is a meaningful expansion of open African-language AI infrastructure. Its success will be measured by whether developers can build reliable services on top of it, and whether speakers of the covered languages find those services useful and accurate.