By Sunbird AI — October 10, 2025
Quick links
Sunflower web interface
Technical report
Models on HuggingFace
Today we are releasing two new open-source LLMs, Sunflower 14B and 32B, which understand 31 Ugandan languages. They are focussed on translation, and are a new state of the art in 24 of these languages. Based on Qwen 3, Sunflower has some ability to answer questions and carry out a range of creative tasks.
These languages represent a cross-section of Africa’s linguistic richness. Sunflower is part of a wider effort to make technology work for everyone, in the languages people actually use.
It was built through collaboration with universities, media houses, and community organizations. Much of the data behind it came from non-digital sources: printed books, school materials, radio programs, and cultural archives. In partnership with Makerere University, Simba FM, the Cross-Cultural Foundation of Uganda, Backup Uganda, and other collaborators, the team digitized and transcribed more than 500 hours of radio and thousands of pages of text. The result is a training dataset and model that reflects how people in Uganda actually speak, learn, and share ideas. Sunflower is not trained on communities, but is built with them.
The accompanying paper, “A New Approach to Expanding Coverage of African Languages in Large Language Models,” explains the datasets, architecture, and evaluation methods.
What this means in practice
Sunflower was not built for novelty; it was built for use.
In classrooms, it enables learning materials to appear in children’s first languages, supporting inclusive and equitable education in line with SDG 4 (Quality Education).
In clinics, it helps nurses and patients understand each other clearly, improving access to information and advancing SDG 3 (Good Health and Well-being).
In governance, it translates public notices and official communication into local languages, strengthening participation and transparency, as outlined in SDG 16 (Peace, Justice and Strong Institutions).
In archives, it turns endangered stories and proverbs into searchable text, helping preserve cultural heritage and promote linguistic diversity in line with SDG 10 (Reduced Inequalities) and SDG 11.4 (Protect the World’s Cultural Heritage).
Sunflower is a public resource, not a commercial product — just as Sunbird itself is a not-for-profit organisation. It is meant to support research, education, translation, and information access and to show that language technology can be built locally with the same efficacy as anywhere else in the world. The models have straightforward Apache 2.0 licensing, and (unlike proprietary models such as ChatGPT) can be adapted and extended by anyone.
Limitations
While the models have strong performance across Ugandan languages, they are still prone to making mistakes. We conducted several rounds of testing with speakers in different languages and identified potential issues:
•Factual recall can be unreliable, and hallucinations may occur.
•The models do not currently perform as well for slang or informal language.
•Some model responses are in the wrong language.
•Glitching can occur, where the model repeats itself or produces unintelligible output.
For practical usage, we recommend validating the model’s outputs for a particular application and fine-tuning as necessary. We are working on improving the models and will be providing updates.
Beyond Translation: Making Language Visible.
For decades, digital systems have spoken the languages of power. Most of the current leading LLMs start with English and add a handful of widely spoken languages, leaving thousands of others, and the people who speak them, out of the digital conversation.
Sunflower takes a different path. It focuses on one of the continent’s most linguistically diverse countries, where over forty languages span three major families: Bantu, Nilotic, and Central Sudanic. By training on them collectively, the model learns to transfer understanding between related languages and produces more reliable, natural responses.
The first release focuses on Uganda, but the approach can scale across the region. The same framework can support tools for East Africa’s 80-plus languages, strengthen multilingual education, and improve communication in public services. Every translation is more than a sentence rendered correctly. It is a bridge repaired between knowledge and access.
It is sometimes said that capitalism assumes accessibility where there is availability, without factoring in inequality. Sunflower challenges that logic. It shows that making technology inclusive requires more than open data or open access; it requires understanding who has been left out and building with them in mind.
It stands as a practical example of what inclusive technology can look like: research grounded in place, open to the world, and built to last.
When language becomes visible, so do the people who speak it.