When Ernest Mwebaze built Sunflower LLM, a language model that can use 31 of Uganda’s languages, he didn’t turn to Google, Microsoft, or OpenAI. He built it on Qwen 3, a Chinese open-source model developed by Alibaba.

It’s a choice shared across the continent. African developers have turned overwhelmingly to Chinese platforms—like DeepSeek, Qwen, and Kimi—to build artificial intelligence models in their own languages. Chinese platforms are faster and cheaper to train, as well as open-source, which is an attractive combination for developers, according to Shikoh Gitau, a leading researcher in African AI and CEO of Qhala.

It also makes them more suited to Africa’s diverse language landscape. UNESCO estimated that there are between 1,500 and 3,000 languages spoken across the continent. Some, like Hausa and Swahili, have tens of millions of speakers across multiple countries. Others, like Kakwa, have a few hundred thousand. Uganda alone has 41 spoken languages.

To train a large language model (LLM), you need data. For languages like English and French, that is easy to find. There are centuries of literature, comprehensive dictionaries, and grammars that an AI can incorporate, as well as an overwhelming amount of modern digital content. But most African languages were not written down before colonization and do not have the volume of data required to train an LLM.

“If models are only available in certain western languages, you’re excluding a lot of people from this technology revolution,” Mwebaze said. “A kid who speaks English and uses ChatGPT is very different from the kid who doesn’t know English and hence can’t use ChatGPT.”

The solution, Gitau said, are small and specialized language models (SLMs and SSLMs) that can be built on minimal data sets and focus on specific applications, like agriculture, health, or education. “Africa is going to win AI on minimum viable intelligence,” she said. Small models serving specific needs, and the best platforms to train them on are currently Chinese.

Kimi, a model developed by Beijing-based Moonshot AI, costs around $3.40 per million output tokens. Anthropic’s Opus 4.7 and OpenAI’s GPT-5.5 cost $25 and $30 per million output tokens respectively.

The disparity can be even worse for African developers. Gitau said that Qhala’s recent research has shown that training a model on an African language can cost three to 30 times more than English.

Gitau called it the “tokenization bias.” “Our languages are hard, they are not documented, they are not digitized, they are not in a format that anybody else is willing to understand them,” making them more difficult, and therefore expensive, to use, she said.

China is not leaving its position in Africa’s AI landscape to chance. In April, the Chinese government launched an AI competition for young African developers. The winners will go to China on study visits. When these top AI developers return home, they will have spent six months learning how to use Chinese models.

But if African AI is built on Chinese foundations, the dependencies that follow may be hard to reverse. “My biggest worry is that we are locked in an ecosystem that doesn’t have policies that you can extract yourself from,” Gitau said.

AI is often described as a leapfrog opportunity for Africa. In the same way that mobile money superseded the need for traditional banks, and mobile phones meant countries did not need to build landline networks, AI has the potential to leapfrog underdeveloped health care, education, and agriculture across the continent.

A2SV, an African AI incubator, launched Skillbridge in 2024. Working in Amharic and Afaan Oromo, it uses AI to help students in Ethiopia study for university entrance exams. It recently expanded to Rwanda.

While Rwanda is the only country thus far to sign a deal to integrate AI into its government, it is only a matter of time before these models are more widely used. In the same way that countries have found themselves reliant on Chinese infrastructure through the Belt and Road Initiative, they will find their AI developments similarly reliant.

The story, Gitau said, has played out before with smartphones. Transsion, a Chinese manufacturer, produces 44 percent of the continent’s smartphones in large part because it made affordable phones marketed solely to Africa. “Everybody now has a Chinese phone, and I feel like that’s what’s happening with AI,” Gitau said. “By the time [Western companies] realize there’s a market, there will be no market.”

To Africa, “there’s not Chinese or non-Chinese ecosystems,” Gitau said, there are just the best tools to build the AI that Africa wants. “Pitching China against Africa is not going to work.”

“There’s a big war between the U.S. and China on AI. We are not going to be part of that war,” Mwebaze said. The United States and China see themselves in a race for AI dominance, but for African developers, it is a choice between technologies, and they will choose whichever one works best for them. Up until now, that has been Qwen, DeepSeek, and Kimi.

The race, as Western companies understand it, has barely begun. Africa remains woefully behind on AI adoption. South Africa has the highest rate of AI usage across the continent, with 23 percent of people saying they have used generative AI. The Democratic Republic of the Congo is at 9 percent. Rwanda, which recently signed a deal with Anthropic to integrate AI into its government and education systems, is only at 7 percent.

A report by Microsoft found that Asian countries have seen rapid growth in AI usage due to increased support for local languages. Since June 2025, South Korea, Thailand, Japan, Mongolia, and Laos have all seen more than 30 percent growth.

For Mwebaze’s new work, he turned to Google’s Gemma. Google’s new smaller models require less compute to run and can be embedded in phones, he said. They can also be trained for speech and text, which is particularly important in areas with lower literacy rates. The cost of training a Gemma model is comparable to Qwen. Mwebaze is, however, among the few people working on Gemma, he said.

There is room to supersede Chinese platforms as AI use grows. Developers will go with whoever offers them the best technology to meet their specific needs of SLMs and SSLMs available in their own languages. Up until now, that has been Qwen, DeepSeek, and Kimi. Their dominance is not guaranteed, but Africa’s future is currently being written in languages that Western companies are not learning to speak.