The Lightweight Stack: African Models and Data in Use, and How They Were Built

The introductory article in this series offered a broader view of what participation in AI can involve, with building frontier models as one route among several. It also introduced early evidence for this view, including the AfriBERTa and InkubaLM language models and the Masakhane research community. This article examines that evidence in greater depth. It considers models, datasets, and working systems produced by African researchers and builders; the methods that made them affordable; and what they suggest for public policy. The examples span agriculture, the environment, health, and language.

Working within the double bind

Much of the work on lightweight AI begins from a shared set of constraints. A 2021 study described them as the “low-resource double bind”: scarce data and scarce computing power occurring together (Ahia et al. 2021). The study examined English-to-Hausa, Igbo, and Yoruba translation models and reduced their size through pruning, a technique that removes the parts of a model that contribute least to its results. The smaller models maintained their performance on frequently occurring sentences, handled rare sentences less well, and coped better with text that differed from their training material. The authors concluded that this approach offers a promising response to the double bind.

The concept matters because it describes the typical working conditions of most African AI researchers, and each approach discussed below can be understood as a response to it.

Models designed to be small

Several African-led teams have shown that working with less data and less computing power can still produce strong models. AfroLM, a language model covering 23 African languages, was trained on a dataset 14 times smaller than those used for comparable models. It nonetheless outperformed several larger multilingual models on tasks that include identifying the names of people, places and organisations in a text, sorting texts by topic, and determining whether a text expresses a positive, negative or neutral view (Dossou et al. 2022).

A different approach adapts an existing model instead of training a new one. A 2022 study adapted a single model to 17 African languages, along with English, French, and Arabic. Before doing so, it removed parts of the model’s vocabulary that served writing systems not used in these languages, reducing the model’s size by about half. The adapted model performed comparably to models adapted to each language separately while requiring far less storage (Alabi et al. 2022).

A third approach attaches adapters, small add-on components, to an existing model and trains only those components. A 2023 study tested adapters for African languages using only free computing resources and found that they performed comparably to large models that require substantial computing power (Leong et al. 2023). This finding matters to governments planning investment in computing infrastructure. National AI strategies often focus on building large computing facilities, and these remain important. The study shows that modest but reliable access to computing, including shared and freely available resources, can also support useful research, thus investment in this kind of access can complement investment in large facilities.

These results are consistent with those reported for AfriBERTa and InkubaLM in the first article (Ogueji et al. 2021; Tonja et al. 2024). Across teams and methods, small models have performed competitively on the tasks they were built for. Work in this area continues. In May 2026, researchers at the University of Cape Town released MzansiLM, described as the first publicly available language model of its kind designed to support all 11 written South African languages, together with the dataset used to train it (The Brief 2026).

Systems in use

Research results matter most when they reach the people they are meant to serve, and several lightweight systems have moved from the lab into use or toward it. The examples that follow show how modest models, built for specific tasks and settings, support the work of farmers, health workers, and public agencies.

Agriculture offers a clear illustration. Cassava is an important crop for many households in East Africa, and two viral diseases, cassava mosaic disease and cassava brown streak disease, are among the greatest causes of cassava losses. Diagnosing these diseases accurately in the field has traditionally depended on agricultural extension officers, who advise farmers on crop management but are often few in number. PlantVillage Nuru was developed to support this work. Built as a public good by Penn State University’s PlantVillage together with the Food and Agriculture Organization and international agricultural research centres, it allows farmers and extension officers to diagnose cassava disease using an ordinary smartphone, without the need for an internet connection. A field evaluation in East Africa compared Nuru’s performance with that of extension officers and farmers. Nuru identified disease symptoms with 65% accuracy, compared with 40 to 58% for extension officers and 18 to 31% for farmers. Its accuracy rose further to 74-88% when users examined six leaves per plant (Mrisho et al. 2020).

Maternal health provides a second example. In Kenya, Jacaranda Health runs PROMPTS, a text-message service that provides pregnant women and new mothers with health information and allows them to ask questions about their symptoms. As the service grew, the volume of incoming questions made it increasingly difficult for its help desk to quickly identify the most urgent cases. By the time the service introduced an AI triage model in June 2022, it had more than 350,000 active users and received about 1.1 million messages a month, many of them written in a mix of English and Swahili (Zhang et al. 2023; Penn State 2023). The triage model reviews these messages and flags those indicating potential danger so the help desk can address them first. In preliminary field tests, it was about 17% more accurate than the earlier model at identifying high-risk conditions, and it reduced the help desk’s workload by about 12% (Zhang et al. 2023). At the time of the study, human agents answered every message, and the model determined the order in which they did so.

Jacaranda Health has since extended this work by building its own Swahili language models. UlizaLlama is based on Meta’s openly available Llama 2 model, which the organisation trained further on 321 million units of Swahili text and equipped with a new Swahili vocabulary of 20,000 word units. The model was then fine-tuned using LoRA, a method that trains a small set of added components while leaving the original model unchanged, keeping adaptation costs low (Jacaranda Health n.d.). PROMPTS now uses a customised language model, UlizaMama, to support its clinical help desk (Agency Fund n.d.), and staff review and audit the quality of the answers it provides. The service now reaches an estimated 700,000 users each year and has expanded to parts of Ghana, Nigeria, and Tanzania, with versions for Twi and Hausa speakers (NPR 2026). In 2024, Meta awarded the organisation a US$300,000 grant to expand PROMPTS in Kenya, Ghana and Eswatini (Meta 2024).

Laboratory diagnosis is a third area where lightweight approaches can help. In many African health systems, diagnosis has long depended on a limited pool of trained staff. A nationwide survey of Ghana’s peripheral laboratories in the early 2000s found that 74% of the 693 laboratory staff were assistants with no professional qualifications (Bates et al. 2004). Research at Makerere University in Uganda shows one way AI could help address this shortage. Researchers there built a low-cost adapter that attaches a smartphone to an ordinary microscope, and then trained image-recognition models to detect malaria parasites in blood samples, tuberculosis bacteria in sputum samples and intestinal parasite eggs in stool samples. The models achieved very high accuracy in all three tasks (Quinn et al. 2016).

Lightweight methods also extend beyond health and agriculture to environmental monitoring. Reference-grade air quality monitors are expensive, which limits how widely they can be deployed in African cities. AirQo, a Makerere University project, responded by designing low-cost air quality sensors for cities in sub-Saharan Africa. Such sensors are affordable, but weather and other surrounding conditions can cause them to overestimate or underestimate pollution levels. To address this, the project developed machine learning models to correct the sensors’ readings and deployed them across a network of more than 120 devices (Adong et al. 2022). The correction allowed a dense network of affordable devices to provide more accurate information for air quality management.

A common feature that runs through the health and agriculture systems described in this section is that in each case, a person makes the final decision. Extension officers and farmers decide how to act on Nuru’s diagnosis, PROMPTS staff respond to flagged messages and audit the service's answers, and in the Makerere design, laboratory staff work with the microscopy models' results. This has practical significance for regulation. These designs show human oversight that already works in practice, and rules governing higher-risk uses of AI can recognise and build on them.

Data and evaluation as participation

Building models is only one way to take part in AI development. Every model depends on the data used to train it and the tests used to judge whether it works, and both shape whose needs a system will meet. Where data reflecting African languages, accents and settings is scarce, models perform poorly for African users, and there is often no reliable way to measure how poorly. Some of the most consequential African contributions to AI therefore take the form of datasets and evaluation tools, which make it possible both to build better systems and to hold existing ones to account.

AfriSpeech-200 illustrates the first of these roles. Doctors in busy African clinics may see more than 30 patients a day, a heavy workload that speech recognition tools could ease by turning spoken notes into written records. Such tools are widely used in health systems elsewhere, but their performance drops considerably for speakers with accents underrepresented in their training data. To address this gap, AfriSpeech creators developed a publicly available benchmark for measuring how well these tools handled African-accented clinical speech by  assembling 200 hours of English speech from 2,463 speakers representing 120 African accents in 13 countries, covering both clinical and general content (Olatunji et al. 2023). The dataset, together with a standard test set and models trained on it, was released publicly so that other researchers and developers can use it to build and evaluate speech tools for African settings.

AfriMed-QA illustrates the second role of evaluation. AI models are increasingly promoted as a way to extend medical expertise in countries facing shortages of doctors and specialists, and many of these models perform well on medical examinations designed elsewhere, such as those used to license doctors in the United States. Strong results on such tests do not establish whether a model gives reliable answers in African clinical settings, where disease patterns, treatment guidelines and health system conditions can differ. AfriMed-QA was created to answer that question. It brings together 15,000 medical questions from more than 60 medical schools across 16 countries, covering 32 specialities (Nimo et al. 2025). It serves as a benchmark, a standard set of test questions against which researchers can compare different AI models. When its creators tested a range of models, they found that performance varied considerably across specialities and countries, and that results on its multiple-choice questions fell below those achieved on the corresponding United States examination questions.

These resources carry two lessons for policy. The first concerns funding, since any developer can use a dataset or benchmark and its value therefore extends well beyond the team that created it, which makes such resources strong candidates for public investment. The second concerns oversight. Benchmarks also give regulators and procurement officers a practical tool for assessing AI systems before they reach the public. Before approving or purchasing a system, these officials can require evidence of how it performs for the population it will serve, measured with tests designed for that population. The variation that AfriMed-QA revealed shows why this matters, since a system that performs well on average, or on tests developed elsewhere, may still perform poorly in a particular country or field.

A new generation building on open models

The methods described so far are increasingly being applied to larger open models, whose trained parameters are released publicly so that others can adapt them. This allows African organisations start with a capable general-purpose model and specialise it for local languages and needs, at a fraction of the cost of developing a comparable model independently.

Sunflower, launched in October 2025 by the Kampala-based non-profit Sunbird AI in partnership with Uganda’s Ministry of ICT and National Guidance, is a prominent recent example. The model covers 31 Ugandan languages and English, and it was built by adapting Alibaba’s openly available Qwen3 models (Sunbird AI 2025a). Sunbird AI reports that, on its own tests, Sunflower outperforms systems from Google and OpenAI in 24 of the 31 languages (PC Tech Magazine 2025). It has also released a compressed version roughly half the size of the original, with little loss of accuracy, lowering the cost of running the model (Sunbird AI 2025b). The partnership with a national ministry also shows how governments can support locally adapted models.

The same approach is now reaching start-ups. In 2025, Meta’s Llama Impact Accelerator selected 40 start-ups in Kenya, Nigeria, Senegal, and South Africa to build products on Meta’s openly available Llama model (Bizcommunity 2025). At the close of the programme, the ventures recognised addressed needs in agriculture, legal services, and health, including a marketplace connecting farmers with buyers in South Africa (Bizcommunity 2025) and an AI assistant for community health workers in Senegal that works without an internet connection (Kajou n.d.). These ventures are early-stage and have yet to demonstrate results at scale. They nonetheless show that open models from large labs allow small teams build tools for specific local problems without the resources needed to develop a model of their own.

What these systems have in common

Although the examples in this article come from different sectors and countries, five approaches recur across them. Together, these approaches explain how African researchers and builders have produced useful systems despite limited data and computing power.

1. The first approach is adapting existing models. The multilingual adaptation study, the adapter experiments, UlizaLlama, Sunflower, and the accelerator’s start-ups all start with an existing model and specialise it for new languages or tasks. This avoids the considerable cost of training a model from the beginning. UlizaLlama’s use of LoRA is closely related to the QLoRA method described in the first article, which reduced the hardware needed for this kind of adaptation to a single specialised processor (Dettmers et al. 2023).
2. The second approach is reducing model size. The examples show several ways to do this such as by pruning parts of a model that contribute least to its results, removing vocabulary the model’s users do not need, and compressing a finished model, as Sunbird AI did with Sunflower. Smaller models cost less to run and can run on more modest equipment.
3. The third approach is using smaller amounts of carefully selected data. AfroLM showed that a model trained on a much smaller dataset can still perform strongly, which matters for languages where large volumes of digital text do not exist.
4. The fourth approach is designing systems to work within the conditions their users face. Nuru and the Senegalese health assistant work on phones without an internet connection, PROMPTS operates over ordinary text messages that do not require a smartphone, and AirQo’s models make inexpensive sensors reliable enough to use. In each case, the system has been shaped around the infrastructure that is actually available.
5. The fifth approach is keeping people responsible for decisions. In the health and agriculture systems described here, the models support the judgement of extension officers, help desk staff and laboratory workers. This has allowed systems with imperfect accuracy to be used with safeguards in place, on the basis that a person can identify and correct errors.

Implications for policymakers

The examples in this article point to several practical steps for governments and regulators.

1. First, treat datasets and benchmarks for African languages and settings as public goods that merit public investment, since their value extends to every developer who uses them.
2. The second is to require evidence of local performance before approving or procuring AI systems, using tests designed for the populations they will serve.
3. The third is to recognise in legal frameworks the forms of human oversight already working in practice, as in the examples of the extension officers, help desk staff and laboratory workers remaining responsible for final decisions.
4. The fourth is to include modest, reliable access to computing within national infrastructure strategies, alongside investment in larger facilities.
5. Finally, partnerships such as that between Sunbird AI and Uganda’s Ministry of ICT show how governments can support the development of locally adapted models.

The limits of the evidence

The models discussed in this article have several limitations. First, their results relate to the specific tasks, languages and conditions under which they were tested, and may not apply elsewhere. The authors evaluated AfroLM and the adapted models described earlier on a defined set of language tasks, and they did not assess performance on other tasks in these studies. Sunbird AI notes that Sunflower may produce inconsistent or less fluent output in some lower-resource languages (Sunbird AI 2025a), and the PROMPTS triage model, although more accurate than its predecessor, does not classify every message correctly.

Second, the evidence on these models varies in age and independence. The cited model studies were published between 2016 and 2025, and several models have since been revised or replaced. For example, UlizaLlama has been followed by a version based on Llama 3, and PROMPTS now relies on a newer model. Many of the reported results come from the developers’ own evaluations, including Sunflower’s comparison with commercial systems, and independent assessments of these models remain limited.

Third, as the first article noted, several of these models are built on open models released by large laboratories. UlizaLlama is based on Meta’s Llama 2, and Sunflower is based on Alibaba’s Qwen3 models. Adapting such models lowers development costs, but some dependence on their developers remains, since licensing and future releases can affect the organisations that build on them.

The next article examines the costs that shape how AI is deployed on the continent, including the energy, language and infrastructure costs of running AI systems, and what they mean for decisions about where and how those systems operate.