UNESCO highlights small AI models for low-resource classrooms
Small, task-specific AI models could help expand access to learning in classrooms with limited electricity, connectivity and devices, according to speakers at UNESCO's Digital Learning Week 2026 in Paris. The models can run on low-cost devices and may be better suited than large language models to clearly defined educational tasks.
- Small models can support bounded tasks such as arithmetic practice and diagnosing misconceptions.
- The Cambridge prototype uses a 2-billion-parameter model on a mid-range phone without internet access.
- UNESCO says task-specific models can reduce energy use by up to 90% for specialized tasks.
- EIDU provides one shared device per classroom in around 14,000 public schools in Kenya and Nigeria.
- Teachers remain responsible for checking and contextualizing AI-generated learning support.
Smaller models can fit specific learning tasks
UNESCO describes frugal AI as tools designed to work with minimal computing power, energy and connectivity. That approach is especially relevant in education systems where classrooms lack reliable electricity, internet access or devices. In Brazil, an average of 29 pupils share each school computer, according to Marcela Lorenzoni, Coordinator of the UNESCO Chair on AIED Unplugged.
Small language models use fewer parameters than large language models and can require less computing power. Rachad Zaki, Director of Cambridge Mathematics at Cambridge University Press & Assessment, said they can support bounded tasks such as arithmetic practice, step-by-step hints and diagnosing specific misconceptions. Large models remain suited to open-ended problems requiring broad knowledge and complex reasoning.
The Cambridge team's prototype maths tutor, (SLM) ², uses a 2-billion-parameter model that runs on a mid-range phone without internet access. The model manages the conversation with the pupil, while a separate calculator checks the maths. UNESCO said its Smarter, Smaller, Stronger report, produced with University College London, found that using smaller, task-specific models for specialized tasks can reduce energy use by up to 90% while maintaining strong performance.
Offline tools are designed for shared devices
Offline-first tools work fully without internet and synchronize data when a connection becomes available. EIDU, an organization based in Germany, provides one shared device per classroom in around 14,000 public schools in Kenya and Nigeria. Teachers use the device to deliver lessons, after which pupils share it for practice. Its AI identifies struggling pupils and, once the teacher confirms, suggests a tailored remedial lesson plan.
In Brazil, an essay-scoring app from AIED Unplugged allows teachers to photograph essays without a connection and upload them for scoring later. The app has reduced the wait for feedback from 4 months to 72 hours. Its design principles include offline operation, support for shared devices and basic digital skills.
EIDU said its broader learning programme has shown gains in foundational literacy and numeracy in randomized evaluations. EIDU and the research organization J-PAL are testing its AI-generated remedial lessons in a large-scale trial in Kenya and Nigeria. The AIED Unplugged team is also making its essay app fully offline.
Teachers and communities remain central
Speakers at the UNESCO event emphasized that teachers must remain involved when AI supports learning. AI systems can make errors, so educators need to check and contextualize AI-generated support. SAM, a Socratic AI maths tutor at Concordia University in Canada, covers a single calculus course and builds on concepts students have already learned. Its developers report an error rate comparable to that of human peer tutors.
UNESCO also highlighted community participation. In Anyama, near Abidjan, Côte d'Ivoire, Impactful Education helps students become creators and peer mentors. Students have created 550 chatbots, animations, games and websites, including projects addressing local issues such as water shortages.
Closing Digital Learning Week 2026, Cédric Wachholz, Chief of UNESCO's Section for Digital Policies and Digital Transformation, said frugal AI is about designing AI around the realities of learners and education systems. The examples presented suggest that for users in resource-constrained classrooms, reliable offline access and a well-defined educational purpose may matter as much as model size.
UNESCO (AI): Small AI models widen access to learning in low-resource settings


