Foundation Models Used in Africa

Earlier articles traced how foundation models came about and set out two features that make them difficult to govern. The first is that their abilities come from training rather than being written in by hand. The second is that one model can sit under many different products (Bommasani et al. 2021).

Both features make the model a product is built on a governance question. If something goes wrong in a model, it can go wrong in everything built on it, so regulators need to know which models are in use, who supplies them, and how they are changed and run. This article sets out what the public record shows about models in use on the continent, in three groups: models built outside Africa, models built elsewhere and reworked here, and models trained in Africa. It then explains how they reach people, what cannot be established, and why that matters.

What counts as a foundation model

"Foundation model" describes how a model is built and used, not how large it is. A model trained on a wide range of data, and able to be put to many different tasks, counts, whether it is one of the largest systems or small enough to run on a laptop. Frontier models are the smaller group at the top end of capability, and the laws written for them identify that group by one measure: the computing used to train the model, counted in arithmetic calculations. California's Transparency in Frontier Artificial Intelligence Act sets the line at more than 10²⁶ operations, which is a 1 followed by 26 zeros. The AI Act treats more than 10²⁵ floating-point operations, or FLOP, as a starting assumption that a general-purpose model carries systemic risk, which a provider can argue against.

These numbers are best read as current markers rather than settled definitions. The amount of computing used in training says little about what a model can actually do, and efficiency gains mean that capability once requiring a very large training run can now be reached with less (Hooker 2024). Analysts who support using compute thresholds treat them as a filter that triggers closer scrutiny rather than a measure of risk in itself, and recommend that any threshold come with a mechanism for revising it as the technology changes (Pistillo et al. 2025). Where the line sits is therefore open. What does not depend on the line is who can train at this scale, since it requires infrastructure held by very few companies (Emberson et al. 2026). No African organisation is known to have trained a model at either level.

That is the global picture. The question for this article is what sits beneath the services people on the continent actually use.

Which models are in use in Africa?

1. Models built outside Africa

The clearest evidence comes from chat assistants. StatCounter, a web analytics company, works out shares from the traffic these products send to the websites it tracks. For Africa in September 2026 it recorded ChatGPT at 73%, Google Gemini at 19%, Perplexity at 4%, Microsoft Copilot at 3%, which runs mainly on OpenAI's models, and Claude at 1% (StatCounter 2026).

These numbers cover web traffic only. They do not account for phone apps, business subscriptions or software that connects to a model directly. Since most internet use in Africa is on a phone, and business and developer use happens outside the browser entirely, the figures describe a slice of use rather than its full extent. They remain the best public sign of which assistants people reach most often, and most of the measured traffic goes to OpenAI's and Google's products.

2. Models adapted in Africa

Beyond direct use, the second group is models built elsewhere and reworked here, usually so they handle African languages.

• Sunflower - Sunbird AI in Uganda built Sunflower 14B and 32B on Alibaba's Qwen 3, covering 31 Ugandan languages and English. The numbers refer to parameters, the internal values a model learns during training, so the 32-billion-parameter version is the more capable of the two and needs more hardware to run. Its builders tested it against other models and found it translated 24 of the 31 languages more accurately (Akera et al. 2025).

• N-ATLaS - Nigeria's government-backed model, launched with Awarri Technologies in September 2025, takes Meta's Llama 3 8B and trains it further on about 392 million tokens of instruction data in Hausa, Igbo, Yoruba and English (Awarri Technologies and Federal Ministry of Communications, Innovation and Digital Economy 2025).

• UlizaLlama - Jacaranda Health in Kenya built UlizaLlama on Llama 2, then extended it using Llama 3 to Hausa, Yoruba, isiXhosa and isiZulu. It uses the model in PROMPTS, its text message service for pregnant women and new mothers (Jacaranda Health 2024).

• Penda Health - Penda Health in Nairobi built a clinical support tool on OpenAI's GPT-4o and published a study of its use across 15 primary care clinics and nearly 40,000 patient visits (OpenAI 2025). The model was built around a clinical task rather than a language.

In each case the service is African while the model under it was built elsewhere.

3. Models trained in Africa

The third group is models trained in Africa, usually developed to address capability gaps in the largest models, particularly in African languages.

Lelapa AI, a South African research and product company, released InkubaLM in 2024. It has 0.4 billion parameters and was trained from scratch on 2.4 billion tokens, the units of text a model learns from, each usually a fragment of a word. Of those, 1.9 billion are in Swahili, Yoruba, Hausa, isiZulu and isiXhosa, with some English and French (Tonja et al. 2024). Lelapa published the model and the datasets behind it under a non-commercial licence, so both are open for research and adaptation but not for commercial use without a separate agreement. Others have since built on them. How widely the model is used is not recorded anywhere public.

A model of this size falls well below both thresholds, as do the models adapted on the continent. Those thresholds are designed to identify a small number of models whose capabilities could create risks at scale, which is a separate question from whether a particular service warrants supervision. A jurisdiction that adopts them acquires a rule that reaches none of the models built or adapted locally. Supervision of services built on those models would require a different trigger, such as where a model is used and what it decides, rather than the computing used to train it.

How these models reach the continent

Knowing which models exist is only part of the picture. How they arrive determines who holds the terms. There are four routes.

• Chat and assistant products, used directly or built into software people already use.

• Direct connections for developers, where a developer's software sends requests to a model running on someone else's computers, as in Penda Health's tool.

• Inside products bought from a supplier, where an organisation buys software and the model within it may never be named.

• Open-weight models, where the trained model itself is published for download, so it can be run or adapted locally. Sunflower, N-ATLaS, UlizaLlama and InkubaLM are all published this way. What "open" allows is the subject of a later article.

All four come with the supplier's terms. The difference with the first three is that the supplier keeps control while the service runs. It decides whether the model remains available, at what price and in which countries, and it can change any of this after an organisation has built on it. Two points follow for regulators. Which countries can buy at all is set by the supplier's own list, which covers most of Africa but leaves out some countries, often because of sanctions (Anthropic 2026). And the processing happens outside the continent, so sending personal data to it engages national rules on transfers abroad.

With a downloaded model, the terms come with the copy rather than the connection. The model runs on the organisation's own computers, so the data can remain in the country and the service does not depend on anyone keeping it switched on. Smaller, compressed versions, like those Sunbird publishes for Sunflower, run on more modest hardware, which puts this route within reach of universities, ministries and small companies. Running a downloaded model locally depends on computing capacity, which is growing on the continent and is itself rented from a small number of operators. A later article examines this in detail.

Two elements continue to rest with the developer that built the model. One is the model itself: the data it learned from, the languages it covers and the choices made in building it all carry into whatever is built on top. The other is the licence, and these differ a great deal. As noted, InkubaLM may not be used commercially. Sunflower is released under Apache 2.0, which allows almost any use. UlizaLlama carries Meta's Llama licence and its rules on acceptable use. N-ATLaS allows use, changes and sharing, but caps it at 1,000 active users in any 30 days, keeps commercial use for a separate agreement, bars uses such as surveillance and military work, and can be withdrawn for breach. Since it is built on Llama 3, Meta's conditions apply as well as Nigeria's.

What cannot be established

The sections above set out what the public record shows. The gaps are as important, because each one forecloses a specific regulatory option.

1. The organisations using closed models. Suppliers do not publish this, and the information is commercially sensitive and often covered by confidentiality agreements with customers. The Penda Health case is known because the company chose to publish it. A regulator cannot establish from public sources which of the banks, hospitals or schools it supervises depend on a particular model. If that model changes, is withdrawn or turns out to be faulty, the regulator learns who is affected only from the supplier or from the organisations themselves.

2. The number of people using the tools. Different counts measure different things. Microsoft's AI Economy Institute counts the share of working-age people who have used one of these tools, putting Kenya at 8.7% in early 2026, South Africa at 23.1%, and Nigeria and Ghana at 10.1% (Microsoft AI Economy Institute 2026). A later survey found 97.5% of Kenyan internet users had used one in the past month (ITWeb Africa 2026). Both are sound, and they count different populations: everyone of working age, and people already online. Neither identifies the model in use, so neither can support a rule that turns on the model, such as requiring that a model used in a public service has been tested in local languages.

3. What is inside bought software. When a bank, insurer or government department buys an AI product, the model inside is usually not disclosed, and sometimes the buyer is not told either. A buyer who does not ask cannot establish where its data goes, which foreign company it now depends on, or what happens if that company changes its terms, and may not be able to answer a regulator or a court asking how a decision was reached. This is the easiest gap to close, since the question can be put into procurement rules and contracts. There are also upcoming guidelines addressing this.

4. More is visible about downloaded models than about the rest. Hosting sites publish download counts, and an adapted model's documentation usually names what it was built from, as Sunflower's, N-ATLaS's and UlizaLlama's do, though the counts are not broken down by country. Any picture drawn from public evidence will therefore show downloaded models more fully than the others, so attention follows the evidence rather than the risk. Downloading raises a further point: anyone can change a model and pass it on without the original builder knowing, so whoever adapts and runs it locally is the party that has to be reachable.

Taken together, regulators in most African countries would supervise how AI is used without a reliable record of what it is built on. They can see the service, the organisation running it, and whatever the supplier chooses to disclose. Problems in the model itself, or in many organisations depending on the same one, would remain out of view.

Why this matters

The gap is structural rather than accidental. In the frameworks drafted so far, the duty to document a model falls on the organisation that built it, while the organisation using it answers for how the service is used rather than for what it is built on. Information about a model therefore passes from developer to vendor to purchaser without reaching a regulator or the public. Three consequences follow for African jurisdictions.

1. There is often no party to ask. Where a model is built abroad and its developer has no local establishment, no entity within the jurisdiction holds the information. The European Union requires providers of general-purpose AI models established outside the Union to appoint an authorised representative within it. In the absence of a comparable mechanism, a regulator depends on what purchasers know and are prepared to disclose.

2. Registers record services, not models. Kenya's Artificial Intelligence Bill, 2026 would establish a public register of high-risk AI systems, recording what is in use and who operates it. It would not record which model each system depends on, so several services relying on the same model would not be identifiable as such.

3. The cost falls due after an incident. As the previous article explained, a fault in a widely used model reaches everything built on it. The first question is which services share that model, and the answer also indicates which evaluations are relevant and which developers to approach. Without a record, each question must be answered individually, after the event.

References

Akera, B., et al. (2025). Sunflower: A New Approach to Expanding Coverage of African Languages in Large Language Models. arXiv:2510.07203.

Anthropic (2026). Supported countries and regions. anthropic.com.

Awarri Technologies and Federal Ministry of Communications, Innovation and Digital Economy (2025). N-ATLaS model card and licence terms. huggingface.co/NCAIR1/N-ATLaS.

Bommasani, R., et al. (2021). On the Opportunities and Risks of Foundation Models. Stanford Center for Research on Foundation Models. arXiv:2108.07258.

California Senate Bill 53 (2025). Transparency in Frontier Artificial Intelligence Act. leginfo.legislature.ca.gov.

Emberson, L., You, J. and Somala, V. (2026). Five hyperscalers now own over two-thirds of global AI compute. Epoch AI. epoch.ai.

Hooker, S. (2024). On the Limitations of Compute Thresholds as a Governance Strategy. arXiv:2407.05694.

ITWeb Africa (2026). Kenya tops global AI usage rankings. itweb.africa.

Jacaranda Health (2024). Jacaranda launches open source LLM in five African languages. jacarandahealth.org.

Kenya (2026). The Artificial Intelligence Bill, 2026 (Senate Bills No. 4 of 2025). parliament.go.ke.

Microsoft AI Economy Institute (2026). Global AI Adoption, Q1 2026. microsoft.com.

OpenAI (2025). Pioneering an AI clinical copilot with Penda Health. openai.com.

Pistillo, M., van Arsdale, S., Heim, L. and Winter, C. (2025). The Role of Compute Thresholds for AI Governance. Institute for Law and AI. law-ai.org.

Regulation (EU) 2024/1689 (Artificial Intelligence Act). eur-lex.europa.eu.

StatCounter (2026). AI chatbot market share, Africa, September 2026. gs.statcounter.com.

Tonja, A. L., et al. (2024). InkubaLM: A Small Language Model for Low-Resource African Languages. arXiv:2408.17024.