A new artificial intelligence (AI) model developed in South Africa could help millions of people use digital tools in their home languages, filling a long-standing gap, local academics say.
A team of researchers at the University of Cape Town (UCT) has developed a language model trained on all 11 official written languages of the country, the first of its kind in South Africa.
The research will be presented at the Language Resources and Evaluation Conference 2026 in Majorca, Spain in May, highlighting the country's growing role in global AI development.
There are two tools at the heart of the project:
- MjansiText, a multilingual dataset; And
- MjansiLM, a language model trained from scratch.
This work was led by Henri Lombard and Dr. Jan Buys together with Dr. François Mayer and a wide team of collaborators.
The study points to language inequality as a major issue in the rise of AI. While AI tools are becoming part of everyday life, they often work best in English and a few widely used languages.
Not many people are involved in this in South Africa. Research shows that only 8.7% of South Africans speak English at home, underscoring the need for tools that work in local languages.
While languages such as isiZulu and isiXhosa have received some attention, others including isiNdebele and Sepedi have been largely ignored. MjansiLM aims to change that
Researchers say the problem is largely due to limited data.
“In language modeling, languages are considered resource-poor, primarily because there are very few and small textual datasets available in these languages for training language models,” Buys said.
He said that while MjansiText is still small compared to the global dataset, it is “larger than previous datasets for South African languages”.
Nine of South Africa's 11 official written languages fall into this “low-resource” category. While languages such as isiZulu and isiXhosa have received some attention, others including izindebele And Sepedi has been largely ignored.
MjansiLM aims to change that. UCT said it is believed to be the first publicly available decoder-only model designed to support all 11 official written languages in one system.
“There have been real advances in language modeling for African languages,” Meyer said, “but most existing models only cover a subset of the languages.”
He said the team's goal was to create a single model focused specifically on South Africa that included all official languages, especially those that are often left out.
The model itself, with 125 million parameters, is relatively small, but the study shows it can still produce robust results.
According to research, it performed well on many tasks and, in some cases, matched or outperformed models more than 10 times its size. For example, on isiXhosa text generation, it achieved a BLEU score of 20.65, competing with much larger systems.
For Lombard, the project began during her master's research on how language models perform in low-resource settings.
“I came to this work through my master's research, which looked at how different language-model architectures perform for low-resource languages,” he said.
Our findings show that the model can work well when properly tailored for specific tasks
— dr jan buys
He noted that most available models only support a few South African languages, and that “MjansiLM was intended to provide a small decoder-only baseline with which future work can be compared and built upon.”
The study also found that the model works best when optimized for specific tasks rather than general use.
“Our findings show that the model can work well for specific tasks,” Buys said, but added that “due to limited training data it is not yet able to work well for general-purpose user interactions or instruction following”.
This helps explain why large AI systems still struggle outside of English, he said.
The researchers emphasized that MjansiLM is not a chatbot But a foundation on which developers can build. “In practice, this means developers can create tools for specific use cases, for example, summarizing information in South African languages or annotating raw data,” Meyer said.
Adopting such a model for concentrated operations could be more effective and economical than using large commercial systems, he said.
The team said the project is only a starting point.
“The progress we have been able to make builds on previously open research from the African natural language processing research community,” Lombard said. He said it is necessary to continue that openness.
Meyer agreed, saying the research community had an important role to play. “This kind of openness often leads to progress,” he said, especially compared to systems where data and methods are not shared.
Times Live