A new five-model open suite adds offline, on-device AI and a speech recognition model that outperforms Gemini, GPT-4o, GPT-5.1, Claude, Meta’s latest, and other closed models on most African languages.
KAMPALA, Uganda, September 24, 2026 — Sunbird AI today released Sunflower v2, expanding its open African-language AI model from 31 Ugandan languages to 67 spanning the continent. The release adds five new open models: a lightweight model that runs offline on mid range smartphones, a speech recognition model covering 51 African languages, and a text-to-speech model covering 20 languages. The five models form an open pipeline running from spoken language to text and back again.
Sunflower’s first release, built on Qwen 3 and covering 31 Ugandan languages, was already outperforming ChatGPT and Gemini 2.5 Pro on translation across two dozen of them. Sunflower v2 keeps that foundation and takes it continent-wide, while adding something the first version couldn’t do: listen and speak, not just read and write.
What’s in the Suite
- Sunflower v2-Qwen3.8-27B, a large language model trained for translation, comprehension, and intent classification across 67 African languages. This model gives improved responses on more complex tasks: on the Sahara LLM benchmark for African languages, it is currently in second place on text generation tasks, just behind Gemini Pro 3, but in front of Gemini Pro 2.5 and all Claude and GPT models.
- Sunflower v2-Qwen3.5-9B: a large language model handling reasoning, translation, and general text tasks across 67 African languages. It has state of the art translation accuracy for African languages, and is small enough to run on a laptop.
- Sunflower v2 -Gemma4-E2B: a compact model built to run directly on a phone and support speech input. It runs entirely on-device with network access. Sunbird is showcasing it through Sunbird Tutor, a proof-of-concept language-learning tool already running on Android in Luganda and Acholi.
- Sunflower v2 ASR: a speech-to-text model covering 51 African languages, outperforming Gemini 3.5 Flash, GPT-4o Transcribe, and Meta’s Omnilingual ASR on most of the languages it was evaluated against. State-of-the-art accuracy on SimbaBench and AfriVox-v2 benchmarks.
- Sunflower v2 TTS: a text-to-speech model covering 20 languages, turning text back into local-language audio.
Open by Default
Every model, and the datasets behind them, is published openly on Hugging Face, free for researchers, developers, and other organizations to use, adapt, and build on.
” Sunflower v2 is a milestone for African languages. It proves that high-precision text and natural conversational voice are achievable right now, and by releasing it as open infrastructure, we’re giving our communities the tools to drive innovation and social impact continent-wide.-Dr. Ernest Mwebaze, Executive Director, Sunbird AI
Access
Sunflower v2 is available now on Hugging Face
Try the sunflower web interface
Try the sunflower speech interface
Access the API.
About Sunbird AI
Sunbird AI is a non-profit building practical AI tools for social impact. Working on artificial intelligence research since 2011 and are based in Kampala, Uganda. Learn more at sunbird-ai.com.
Media Contact
Name: Nimpamya Janat Namara
Email: info@sunbird.ai