Introducing luganda-gemma-1B-it

Most AI tools assume you speak English. For over 7 million Luganda speakers in Uganda, that assumption has meant exclusion from civic information systems, digital health tools, and the platforms that increasingly mediate access to public services.
Today, we are changing that. Amplified Access is releasing luganda-gemma-1B-it, a publicly available AI model purpose-built for English-to-Luganda translation and Luganda conversational AI. It is now available for free on HuggingFace, Kaggle, and Ollama.

Luganda is the primary language of Kampala and central Uganda. It is the language of community meetings, of health consultations, of the conversations that happen at trading centres and local council offices across the country. It is the language in which civic life actually happens for millions of people.
And yet when you ask today's most capable AI systems to speak it, most cannot. They switch to English or produce something no native speaker would recognise. The tools built to make information more accessible have, in practice, made it less accessible for Luganda speakers. They were built without Luganda speakers in mind.
This model is a direct response to that gap. We built it to support the translation of civic content (health advisories, government service information, community rights guidance) into a language that communities can actually use. Not as an add-on. Not as a language pack bolted onto a system designed for English. As the primary focus.

luganda-gemma-1B-it supports English-to-Luganda translation, Luganda-to-English translation, Luganda conversational AI, and language learning tools. It was trained on human-verified content covering agriculture, health, and society: the domains where accurate translation in Luganda is most urgently needed.
A community officer can now use it to translate service information before a community meeting. A health organisation can use it to localise public health guidance. A legal aid team can use it to make rights information legible to people who have never had access to it in their own language. Developers building civic tools for Uganda can integrate it directly into their applications.
It runs locally. No internet connection is required at inference time. No API fees. No data leaves the user's device. For the communities we serve, many of which have intermittent connectivity and real data privacy concerns, this is not a minor feature. It is the difference between a tool that can be deployed and one that cannot.



AVAILABLE NOW
HuggingFace: AmplifiedAccess/Luganda-gemma-1b-it
Kaggle: amplifiedaccess/luganda-gemma-1b-it
Ollama: amplifiedaccessorg/Luganda-gemma-1b-it


How we built it, and why it was possible
One of the persistent myths about African language AI is that it requires massive resources: data centres, large research teams, millions of dollars. We want to be direct about what it actually took: a consumer GPU, a focused dataset of human-verified Luganda translation pairs, and a training approach called QLoRA that makes it possible to teach a large base model a new language without rebuilding it from scratch.
The result is a 52 MB adapter that sits on top of Google's Gemma 3 1B base model. It contains everything we added. The whole fine-tuned model runs on hardware that a small organisation can afford.
On English-to-Luganda translation, it scores BLEU 13.85, up from 0.06 on the untuned base model. That 230-fold improvement comes from human-verified training data: sentence pairs reviewed by native Luganda speakers, not generated by another machine translation system. The quality of the data is what makes the difference, not the scale of the infrastructure.
This matters beyond Luganda. We started here because it is a problem in our neighbourhood. But the same approach (small model, human-verified data, community-relevant domains) applies to Runyankole, Kinyarwanda, Acholi, and dozens of other languages across East Africa that mainstream AI has not reached. luganda-gemma-1B-It is the first model in a programme, not a one-off.

We are expanding the training corpus, running a full benchmark evaluation, and beginning structured human evaluation with native Luganda speakers. BLEU scores are useful technical proxies, but the real measure of this model is whether it helps a health worker in Kampala communicate more effectively with a patient or helps a community organiser explain a land rights process to people who need that information in Luganda.
That evaluation is ongoing. We will share what we find.
And we are already working on the next language.
If you are building civic tools for East African communities, we want to hear from you. If you want to contribute Luganda training data or participate in human evaluation, reach out. If you just want to try the model, it is available now.
The infrastructure for African language AI does not have to be built by someone else, somewhere else, for communities that are treated as an afterthought. We are building it here, with and for the communities that need it.