Building African-language AI is easier than finding the data

The people who will feel the ripple of this story are the millions of Africans who speak ChiShona, Hausa or Nigerian Pidgin every day. They have long been excluded from the AI boom because their languages were invisible in the data that powers modern models. Vambo AI’s launch of MORENA shows that the biggest hurdle is not computing power but the sheer lack of authentic text in these tongues. If this project succeeds, it could finally give a voice to the continent’s linguistic diversity and unlock a new wave of local applications.

Vambo’s co‑founder and CTO, Isheanesu Misi, admits that most of the 65.6 billion African‑language tokens came from synthetic generation. The scarcity of real‑world text meant that the company had to fill the gap with artificially created sentences. This approach is not without risk. “You would find that some of the grammar of the outputs is close but imperfect,” Misi said, pointing out that synthetic data can introduce subtle errors that bleed into downstream applications. Yet, synthetic data also allowed Vambo to steer the model toward specific capabilities, such as tool calling, that might otherwise be missed in a purely data‑driven approach. The trade‑off is clear: speed and breadth versus fidelity and nuance.

Rather than fine‑tuning a general‑purpose foundation model, Vambo built MORENA from scratch. This decision was driven by the observation that legacy models often inherit biases and blind spots when applied to African languages. By starting fresh, Vambo could design tokenisers that use 1.39 times fewer tokens than Gemma 3 and 1.53 times fewer than Llama 3.2 for the same content. The result? A lower bits‑per‑byte score of 1.408 across 12 languages, the best among 26 models tested, including the 8‑billion‑parameter Lugha‑Llama‑8B. This performance gap underscores that a bespoke approach can outperform generic alternatives, especially when the target data is sparse.

The shortage is most acute for languages like isiNdebele and Nigerian Pidgin, which have millions of speakers but almost no AI resources. Misi warned that without adequate data, AI applications will remain “extremely important and necessary, but they can’t be fully….” The sentence trails off, but the implication is stark: without robust data pipelines, the continent’s digital future risks being built on a brittle foundation. If Vambo’s model can inspire other African‑centric AI projects, it could spark a virtuous cycle of data creation, model improvement, and product innovation.

Given the trajectory of African‑language AI, I predict that by 2028, a handful of locally developed models will surpass global competitors in handling African languages. The key will be community‑driven data curation, where speakers contribute to corpora and validate outputs. Vambo’s success with synthetic data is a warning that quality must be monitored, but it also demonstrates that the tools exist to bridge the gap.

The next step for Vambo is clear: open‑source MORENA’s tokeniser and training scripts so that other developers can replicate and extend the work. If they do, the continent will finally have a suite of AI tools that truly reflect its linguistic richness, and the digital divide will begin to close.