UCT Researchers Develop Multilingual AI Language Model

Why It Matters

MzansiLM provides a foundational, open‑source tool that can reduce reliance on costly proprietary models and accelerate AI‑driven services in under‑served South African languages, fostering digital inclusion and local innovation.

Key Takeaways

• MzansiLM supports all 11 South African official written languages
• Dataset MzansiText is the largest publicly available SA language corpus
• Model has 125 million parameters, smaller than commercial LLMs
• Outperforms larger open‑source models on SA language benchmarks
• Released publicly for developers to fine‑tune for local applications

Pulse Analysis

The rise of large language models has transformed how information is accessed worldwide, yet the benefits remain unevenly distributed. In South Africa, eleven official written languages compete for representation, and most global AI services struggle with low‑resource tongues such as isiNdebele or Sepedi. Recognising this gap, a research team at the University of Cape Town set out to create a model that treats all national languages as first‑class citizens rather than afterthoughts. Their effort underscores a broader shift toward inclusive AI, where linguistic diversity is built into the core of model development rather than tacked on later.

The project produced two linked assets: MzansiText, a curated corpus that aggregates publicly available texts across all eleven languages, and MzansiLM, a 125‑million‑parameter decoder‑only model trained from scratch on that data. Despite its modest size compared with commercial giants, MzansiLM achieved higher accuracy on targeted benchmarks for languages like isiXhosa, Sesotho and Tswana, outpacing larger open‑source alternatives. The researchers attribute this success to the focused training regime and the relative homogeneity of the dataset, demonstrating that well‑designed low‑resource models can compete with broader, data‑hungry systems.

By releasing both the dataset and the model under an open licence, UCT invites developers, startups and government agencies to fine‑tune the baseline for specific use cases such as document summarisation, sentiment analysis or public‑service chatbots. This democratizes access to AI in local languages, potentially lowering costs for organisations that would otherwise rely on expensive proprietary APIs. Moreover, the initiative creates a benchmark for future African language research, encouraging collaboration across universities and industry to expand training data and improve model robustness. As more applications emerge, the economic and social impact of native‑language AI could become a catalyst for digital inclusion across the continent.