Public conversation about artificial intelligence often measures progress by size. The models that make headlines are trained on enormous datasets using thousands of specialised chips, at costs that only a handful of companies and governments can meet. When African participation in AI is discussed, it is usually measured against that standard, and the continent is found to fall short on several counts: limited compute, scarce capital, too little usable data and few data centres. The usual conclusion is that Africa will mainly consume AI built elsewhere.
This series, The Lightweight Stack, takes a broader view of what participation can involve. Building frontier models is one route, and it remains open to the continent. Meaningful participation also takes place through smaller and mid-scale models, through adapting open models to local languages and tasks, through building data and evaluation resources, and through developing the institutions and skills that allow AI to spread across an economy. Each of these is a real form of capability, and each is within reach of African researchers, companies and governments today.
Choosing the Right Benchmarks for African AI Development
The assumption that bigger models are always better has also been questioned. Gaël Varoquaux, Sasha Luccioni and Meredith Whittaker challenge two assumptions underlying the scale-focused paradigm: that improved performance is a product of increased scale, and that all interesting problems addressed by AI require large-scale models (Varoquaux et al. 2025). They also argue that this focus sidelines important applications such as health, education and the climate, and concentrates decision-making power in the hands of a few actors. For African policymakers, these neglected applications are often the most important ones, which makes it useful to measure AI participation in ways that extend beyond frontier capability.
Other researchers have begun to describe what an alternative could look like. Roy Bendor, Dave Murray-Rust and Rainer Rehak propose "Small AI" as a counterweight to the bigger-is-better trend. They extend the idea of smallness beyond parameter counts to how locally rooted computation is and whose language and culture a model serves (Bendor et al. 2026). They also note that a model small enough to run on a laptop or other device can be used locally, which eases many concerns such as data sharing. This point carries particular weight in African markets, where data protection and dependence on foreign providers are live policy concerns.
The same assumption shapes regulation. Compute thresholds are the mechanism used in the EU AI Act and in earlier US policy to identify the riskiest models. Sara Hooker concludes that such thresholds, as currently implemented, may not be an effective way to manage risk, because the relationship between compute and risk is highly uncertain and changing quickly (Hooker 2024). African regulators who adopt frameworks designed around the largest systems may therefore overlook the smaller models that are most likely to be deployed in their own markets that require different regulatory oversight.
Evidence from Africa
African researchers have also been demonstrating the value of smaller models for several years. In their 2021 paper challenging the assumptions that language models need large datasets and that low-resource languages depend on high-resource ones to perform well, Kelechi Ogueji, Yuxin Zhu and Jimmy Lin introduced AfriBERTa, a multilingual language model trained only on African languages. They showed that competitive multilingual models can be trained on less than 1 GB of text (Ogueji et al. 2021). AfriBERTa covers 11 African languages and was the first language model for four of them. It was tested on two tasks: picking out the names of people, places and organisations in a text, and sorting texts by topic. It outperformed larger established multilingual models in several languages and was competitive overall.
In 2024, Lelapa AI and its research partners released InkubaLM, a 0.4 billion-parameter model built for isiZulu, Yoruba, Hausa, Swahili and isiXhosa (Tonja et al. 2024). Its developers report that it performs comparably to models with far more parameters and training data on machine translation and question answering. They report the same for two tests drawn from IrokoBench, a benchmark designed to measure how well AI models handle 17 African languages, with test material translated by human translators (Adelani et al. 2025). IrokoBench is made up of three tests. The two used to evaluate InkubaLM are AfriMMLU, which checks general knowledge through multiple-choice questions, and AfriXNLI, which checks whether a model can tell if one statement follows from, contradicts or is neutral towards another. The third test covers mathematical reasoning. InkubaLM also outperforms many larger models on sentiment analysis, which means identifying whether a text expresses a positive, negative or neutral view.
Participation also extends beyond building models as seen in initiatives such as Masakhane, a grassroots research community founded in 2019 whose name means "we build together" in isiZulu. It brings together researchers and volunteers from across the continent to strengthen natural language processing for African languages. It’s machine translation work made the case that a shortage of text is only part of why these languages are poorly served. Who does the research, and whether existing texts can be found and used, matter as well (Nekoto et al. 2020). In their view, the shortage of digital text in a language reflects wider problems in the societies where it is spoken. One example they give is colonial-era suppression of writing in African languages, which reduced the number of books and other publications in those languages. Others are limited access to university education, which means fewer researchers can work on their own languages, and limited access to suitable keyboards and technology for creating digital content. Masakhane's participatory approach produced new translation datasets and benchmarks for over 30 languages, and enabled participants without formal training to make scientific contributions.
Why lighter approaches are a plausible path for Africa
Lighter approaches have become more practical because technical advances have lowered the cost of adapting existing models. Adapting a model, often called fine-tuning, means giving an existing model further training so that it performs better on a particular task or language. A method called QLoRA, published in 2023, makes this far cheaper. It stores the original model in a compressed form that takes up much less computer memory, keeps that model fixed, and trains only a small set of added components. Its developers showed that this reduces memory use enough to fine-tune a model with 65 billion parameters on a single specialised processor (a GPU) with 48 GB of memory, while matching the results of standard fine-tuning (Dettmers et al. 2023). This means an African lab or startup can adapt a capable open model without building its own large cluster of computers for training.
The cost of running models matters as much as the cost of building them. Alexandra Sasha Luccioni, Yacine Jernite and Emma Strubell measured the energy used and carbon emitted when different kinds of models carry out the same tasks. They found that general-purpose models designed to generate text or images used far more energy than systems built for a single task, often by a factor of ten or more, and that this held even when the models being compared were of similar size (Luccioni et al. 2024). In countries where electricity supply is constrained and expensive, that difference directly affects which AI services are viable.
Language adds a further cost. Before a model can process text, a component called a tokeniser breaks the text into small units known as tokens. A 2023 study found that the same text translated into different languages can produce up to 15 times as many tokens, depending on the language (Petrov et al. 2023). Since commercial AI services usually charge by the token, this leads to unequal costs for using them. It also means longer processing times, and less text in some languages fits within the amount a model can take in at once. Where African languages are affected in this way, their speakers can end up paying more and waiting longer when using models designed primarily around English. This strengthens the case for models built with those languages in mind.
Participation as diffusion
A further argument for this series comes from economic history. In Technology and the Rise of Great Powers, Jeffrey Ding studies general-purpose technologies, such as electricity and computers, which are used across almost every sector of an economy. He asks why some countries were better than others at spreading these technologies widely, a process known as diffusion, and gives less weight to which country invented them first. Drawing on the industrial revolutions in Britain, the United States, Germany and Japan, he finds that the countries that gained most were those whose institutions helped new technologies spread throughout the whole economy (Ding 2024). Ding's study concerns competition between great powers, so it applies to African economies by analogy. Its central lesson still holds: the economic gains from a general-purpose technology depend on how widely and effectively an economy uses it.
Development institutions have reached similar conclusions. UNCTAD's Technology and Innovation Report 2025 identifies infrastructure, data and skills as the three areas developing countries most need to strengthen if they are to benefit from AI (UNCTAD 2025). The African Union's Continental AI Strategy, sets out an approach to AI that puts African priorities and development needs at its centre (African Union 2024). A lightweight approach offers one practical route to delivering on these commitments. Smaller models need less infrastructure, depend on local data, and can be adapted by African researchers and companies, which builds skills in the process.
On the limits of the lightweight approach
This series will also be candid about the limits of the lightweight approach. The first is dependence. Many small, adapted models are built on open-weight models released by large laboratories, so lightweight participation often still relies on the work of frontier developers. The second concerns how far the evidence reaches. Results such as those reported for InkubaLM come from the developers' own tests on specific tasks and languages (Tonja et al. 2024). They show that small models can work well for targeted uses, but they do not show that small models match the largest systems across the board.
The third limit concerns who is making the case. Some prominent advocacy for small models comes from companies with commercial interests in the hardware used to run them. A 2025 position paper by researchers at NVIDIA, whose chips are used to run AI models of all sizes, argued that small models will become the standard for many automated AI systems (Belcak et al. 2025). Arguments like this can still be correct, but should be weighed carefully.
A smaller model is also not necessarily a safer one. A legal analysis of compute thresholds points out that some risks come from small, specialised models (Pistillo et al. 2025). Its example is a drug discovery tool that researchers repurposed to look for the most toxic molecules instead of the least toxic ones. In under six hours it proposed tens of thousands of candidates, including known chemical warfare agents and new compounds predicted to be just as deadly (Urbina et al. 2022). Rules that rely only on model size would not catch tools like this.
Finally, the environmental benefits of small models may be smaller than they first appear. Advocates of "Small AI" warn of a rebound effect that would happen when small models make AI cheaper and more energy-efficient to run, people may use it much more, and the extra use can cancel out the energy savings (Bendor et al. 2026). Engaging openly with these limits is part of making the case for lightweight approaches properly.
What the series will cover
Over the coming instalments, The Lightweight Stack will examine why scale is not the only measure of progress; African models and datasets that are already in use; the techniques that make adaptation affordable; the energy and language costs of different deployment choices; and the policy and regulatory frameworks that would support a lightweight approach across the continent. The aim throughout is to give policymakers, lawyers and practitioners a clear picture of the forms of AI participation available now, and what it would take to expand them.
References
Adelani, David Ifeoluwa, Jessica Ojo, Israel Abebe Azime, et al. (2025), IrokoBench: A New Benchmark for African Languages in the Age of Large Language Models, Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), 2732–2757, Albuquerque, NM: Association for Computational Linguistics.
African Union (2024), Continental Artificial Intelligence Strategy: Harnessing AI for Africa's Development and Prosperity, Addis Ababa: African Union.
Belcak, Peter, Greg Heinrich, Shizhe Diao, et al. (2025), Small Language Models are the Future of Agentic AI, arXiv preprint, 2506.02153.
Bendor, Roy, Dave Murray-Rust, and Rainer Rehak (2026), Small AI: A Degrowth Imaginary for Designing With/For Artificial Intelligence, DRS2026: Edinburgh, DRS Biennial Conference Series, Design Research Society.
Dettmers, Tim, Artidoro Pagnoni, Ari Holtzman, et al. (2023), QLoRA: Efficient Finetuning of Quantized LLMs, Advances in Neural Information Processing Systems, 36, 10088–10115, Red Hook, NY: Curran Associates.
Ding, Jeffrey (2024), Technology and the Rise of Great Powers: How Diffusion Shapes Economic Competition, Princeton Studies in International History and Politics, Princeton, NJ: Princeton University Press.
Hooker, Sara (2024), On the Limitations of Compute Thresholds as a Governance Strategy, arXiv preprint, 2407.05694.
Luccioni, Alexandra Sasha, Yacine Jernite, and Emma Strubell (2024), Power Hungry Processing: Watts Driving the Cost of AI Deployment?, Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency, 85–99, New York: Association for Computing Machinery.
Nekoto, Wilhelmina, Vukosi Marivate, Tshinondiwa Matsila, et al. (2020), Participatory Research for Low-resourced Machine Translation: A Case Study in African Languages, Findings of the Association for Computational Linguistics: EMNLP 2020, 2144–2160, Association for Computational Linguistics.
Ogueji, Kelechi, Yuxin Zhu, and Jimmy Lin (2021), Small Data? No Problem! Exploring the Viability of Pretrained Multilingual Language Models for Low-resourced Languages, Proceedings of the 1st Workshop on Multilingual Representation Learning, 116–126, Punta Cana: Association for Computational Linguistics.
Petrov, Aleksandar, Emanuele La Malfa, Philip Torr, et al. (2023), Language Model Tokenizers Introduce Unfairness Between Languages, Advances in Neural Information Processing Systems, 36, 36963–36990, Red Hook, NY: Curran Associates.
Pistillo, Matteo, Suzanne Van Arsdale, Lennart Heim, et al. (2025), The Role of Compute Thresholds for AI Governance, George Washington Journal of Law & Technology, 1, 26ff.
Tonja, Atnafu Lambebo, Bonaventure F. P. Dossou, Jessica Ojo, et al. (2024), InkubaLM: A Small Language Model for Low-Resource African Languages, arXiv preprint, 2408.17024.
UNCTAD (2025), Technology and Innovation Report 2025: Inclusive Artificial Intelligence for Development, Geneva: United Nations.
Urbina, Fabio, Filippa Lentzos, Cédric Invernizzi, et al. (2022), Dual Use of Artificial-Intelligence-Powered Drug Discovery, Nature Machine Intelligence, 4, 3, 189–191.
Varoquaux, Gaël, Sasha Luccioni, and Meredith Whittaker (2025), Hype, Sustainability, and the Price of the Bigger-is-Better Paradigm in AI, Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency, 61–75, New York: Association for Computing Machinery.
