Small AI Models Conquer the Real World

What Happened?

In 2019, Adebayo Alonge, founder of a Nigerian startup, was in a Cape Town hotel ready to demonstrate RxScanner, a portable device that uses infrared spectroscopy and a cloud-based AI model to detect counterfeit drugs. This problem is critical in Africa, where the World Health Organization estimates that one in ten medical products in developing countries is falsified or substandard, causing thousands of deaths annually. However, during the demo, the lack of internet connectivity rendered the device inoperable. In just two hours, his team managed to compress the AI model to run entirely on an Android phone offline. That incident gave rise to an offline version of the product now used in over a dozen countries, including Ghana, Kenya, Myanmar, and Nigeria. This case, reported by IEEE Spectrum and circulated by Slashdot, illustrates the rise of small AI—models optimized to run on low-power devices without relying on the cloud.

Other notable examples include the work of Bala Murugan and his team at the Vellore Institute of Technology (India), who developed a drone capable of photographing cashew plants and detecting disease-indicating spots by processing images locally, without external computers or internet. In Uruguay, similar systems have been implemented to identify ant infestations in vineyards. In Brazil, researchers have used small AI to perform electrocardiograms from an Arduino device, bringing cardiac diagnostics to areas without access to complex medical equipment. Marcelo Jose Rovai created a TinyML model that, running on just 3 watts of power, detects puddles where mosquitoes might breed, helping prevent diseases like malaria. These applications demonstrate that AI does not necessarily require large data centers; it can be local, efficient, and accessible.

Why Is It Important?

The move toward small AI represents a paradigm shift from the dominant trend of massive models like GPT-4 or Gemini, which require multi-billion-dollar data centers and consume enormous amounts of energy. According to a study by the University of Massachusetts Amherst, training a single large model can emit as much CO2 as five cars over their lifetimes. In contrast, small AI offers key advantages in accessibility and resilience. By running locally, it eliminates dependence on the internet, making it ideal for regions with limited infrastructure, such as rural areas in Africa, Asia, or Latin America. It also reduces latency, improves privacy by processing data on the device, and lowers operational costs by avoiding cloud subscription fees.

Moreover, small AI fosters local innovation. Developers can create solutions tailored to specific problems without requiring large investments in hardware or connectivity. This is particularly relevant in sectors like precision agriculture, where farmers need affordable tools to detect crop diseases, or public health, where early detection of diseases like malaria can save lives. The ability to run on batteries and solar panels further extends its reach in remote environments.

What Consequences Will It Have?

The adoption of small AI could accelerate digitalization in emerging economies where telecommunications infrastructure is limited. According to the International Telecommunication Union, 37% of the world's population still does not use the internet, and in sub-Saharan Africa the figure exceeds 60%. Small AI allows communities without cloud access to benefit from artificial intelligence. Sectors like agriculture, healthcare, and environmental monitoring will be the first to transform. For example, drones with local AI can help farmers optimize pesticide and fertilizer use, reducing costs and environmental damage. In healthcare, devices like RxScanner can verify drug authenticity in rural pharmacies, combating a problem that the WHO estimates affects 10% of products in developing countries.

In the long term, this trend could decentralize artificial intelligence, reducing dependence on a few tech giants that control large models (such as OpenAI, Google, or Microsoft). Small AI fosters a more diverse ecosystem where startups and local developers can compete with specialized solutions. However, challenges exist. Small models are often limited in generalization capability and require very specific datasets to be effective. Additionally, maintaining and updating these distributed models can be complex, especially if devices lack connectivity. Despite this, advances in model compression (such as quantization and pruning), low-power hardware (like microcontrollers and FPGAs), and federated learning techniques are rapidly improving the balance between performance and efficiency.

What Should Readers Know?

Readers should understand that AI is not just about large cloud models. Small AI is already saving lives, protecting crops, and improving medical diagnostics in resource-limited settings. For businesses and developers, this trend opens opportunities to create more accessible and sustainable products. Moreover, it is a reminder that technological innovation does not always require cutting-edge hardware; sometimes the most effective solution is the simplest and most local. Small AI also raises questions about privacy and security: by processing data locally, the risk of breaches is reduced, but updating models on offline devices remains a challenge. In any case, the future of AI lies not only in large data centers but also in the small devices we carry in our pockets or that fly over crop fields.

"The future of AI lies not only in large data centers but also in the small devices we carry in our pockets or that fly over crop fields."