South African startup builds AI model for 12 African languages

If you use an AI chatbot in South Africa, there is a good chance you are better off asking it a question in English. That is less useful for a country where millions of people use African languages every day. South African startup Vambo AI has released MORENA, a 1.5-billion-parameter language model trained from scratch for 12 African languages. It also supports English and French.

The company has taken a different route from the large general-purpose models dominating the AI market. Instead of trying to build a model that knows a little about everything, Vambo has concentrated on languages that have not received the same attention from AI developers.

That is a fairly specific problem to work on, but it has a practical use.

A company cannot build much of a customer-service operation around AI if the system does not understand the language its customers use.

African languages have been an awkward fit for AI

Most of the large language models used today were trained on enormous amounts of online material. English has an obvious advantage here. There is a huge amount of English-language material available to train models on. Many African languages have far less digital material. That affects how well AI systems understand them. It can show up in simple things: a question that is misunderstood, a poor translation, or a response that sounds unnatural. For a business, those problems are not academic.

If a bank puts a chatbot in front of its customers, the chatbot has to understand the questions customers actually ask. If a retailer wants customers to search for products in their own language, the system has to understand those searches. Vambo has built MORENA around 12 African languages rather than treating them as an afterthought.

Vambo is not chasing the biggest model

MORENA has 1.5 billion parameters. That is small compared with the biggest models being built by the world's largest technology companies. But Vambo is not competing on model size. The company says MORENA was trained from scratch for its target languages and can be adapted by startups and researchers. That could make it useful to developers who need a model for a particular application rather than a giant system designed to handle almost any request.

For example, a company building a customer-support product may need a model that understands isiZulu particularly well. It may have little reason to pay for or run a much larger model built for thousands of unrelated tasks. Whether MORENA can deliver that level of performance in actual products will depend on how developers use and test it.

The customer may never see the AI

If this works, most people using products built with MORENA may never know it is there. They could be talking to a bank through a messaging service, asking a retailer about a product or using a voice service. The important part for them would simply be whether the system understands what they said. That is where African-language AI has a fairly straightforward commercial case.

Businesses want to move more customer interactions online, but forcing every customer to communicate in English limits who can comfortably use those services. A model that handles local languages well could sit underneath those services. But someone still has to build the products. Vambo has released the underlying model. A bank, retailer or software company would still need to connect it to its own systems and make sure the responses are accurate.

The hard part starts after the launch

There is a lot of attention around AI models when they are released. For Vambo, the less glamorous work comes next. Developers have to try MORENA in real applications. Businesses have to decide whether it performs well enough to use. And the model has to handle the different ways people actually speak and write the languages it supports.

A language can look very different in a training dataset from the way it appears in a WhatsApp message, a customer complaint or a voice recording. That is where the claims around African-language AI will eventually be tested. Vambo has built a model around 12 African languages. Now it needs people to build things with it. If they do, the significance will not be that Africa has produced another AI model. It will be that businesses can start making digital services that do not assume every customer is most comfortable in English.