RxScanner, and more: how small AI models can save lives and do much more

Adebayo Alonge found himself in a hotel room in Cape Town in 2019, ready to show his most important audience how RxScanner could detect counterfeit drugs that kill thousands of people every year in Africa. That morning, the system crashed: the connection to the US data center, over 14,000 kilometers away, failed right in front of the stakeholders he needed to convince.

RxScanner is a portable spectrometer that analyzes a pill with infrared light and sends its molecular profile to an AI model linked to a pharmaceutical database. In its original version, however, the response took more than five minutes and relied on a stable connection, a luxury not guaranteed in many areas of the continent.

Alonge asked his engineers to reduce the model to a lightweight, offline version capable of running entirely on an Android phone. The team delivered it in just two hours, saving the demo and giving rise to a device that today authenticates drugs even without broadband, computers, or reliable electricity, and is already in use in over a dozen countries including Ghana, Kenya, Myanmar, and the founder's home country of Nigeria.

From Indian Drones to Anti-Malaria Sensors

The case of RxAll is not isolated. At the Vellore Institute of Technology in India, researcher Bala Murugan and his colleagues have developed a drone that photographs cashew trees and autonomously recognizes typical spots of fungal diseases. All processing occurs onboard the aircraft, without the need for a computer, connection, or remote server.

In a Uruguayan vineyard, a similar compact AI system has been trained to recognize ant infestations, while in various countries, similarly streamlined models are employed to detect the presence of malaria-carrying mosquitoes, a useful task especially where health teams lack access to costly laboratories or reliable data networks.

In Brazil, Professor Marcelo José Rovai from the Federal University of Itajubá has applied the same logic to cardiology: a TinyML model generates electrocardiograms in a lab with a simulated patient, running on a basic Arduino in areas lacking complex medical equipment. His latest experiment uses an Arduino UNO Q board with a Qualcomm chipset, costing about $50: it runs a local language model, collects data from sensors, and analyzes it to identify water pools where mosquitoes lay their eggs, consuming only 3 watts.

For professionals, the term "small AI" refers to models with at most a few billion parameters, against the trillions of more advanced systems. Google DeepMind and Alibaba have fueled this trend with the release of Gemma 4 and Qwen 3.5, both open-weight and adaptable to limited hardware. The World Bank openly supports these projects with funding, mentorship, and policy guidelines: President Ajay Banga reminded Davos that only 0.7% of internet users in the poorest countries have ever used ChatGPT, a statistic that alone explains the urgency for solutions that do not rely on the cloud.

Industry estimates indicate that 33% of smartphones shipped in 2025 will already be capable of running small AI models locally, a share expected to rise to 45% in 2026 and over 51% by the end of 2027. For Alonge, this trajectory is not a fallback but a structural choice: "If no one subsidizes it, most people won't be able to afford those huge models," he explains, adding a more radical belief: "I believe the future of AI is not one large model, centralized in one place." In short, we are increasingly paying attention to small models with a few billion parameters, as they can increasingly make a difference.