UNESCO rewards AI models that do more with less energy 🌱🤖

Cliquez ici pour lire en français

More than 149 participants from 30 countries took part in the Resilient AI Challenge, an international competition focused on building more energy-efficient AI models. Backed by UNESCO, the governments of France and India, and the Sustainable AI Coalition, the initiative recognized three teams that managed to cut their models’ computing needs while keeping performance intact.

Cutting compute without cutting corners ⚡

More specifically, participants worked on leaner versions of open-source and open-weight models supplied by the challenge’s technology partners, including Google, Mistral AI, and Sarvam AI. The goal wasn’t just to shrink the models. Teams had to strike a balance between three core variables: accuracy, system performance, and energy consumption.

To get there, they tackled three concrete AI applications: text generation, audio transcription, and image captioning.

The challenge builds on findings from a joint study by UNESCO and University College London (UCL), titled Smarter, Smaller, Stronger: Resource-efficient AI and the Future of Digital Transformation. The research shows that more deliberate AI system design can significantly cut resource consumption without necessarily hurting performance. The Resilient AI Challenge turned those research findings into working applications.

Three winning teams 🏆

Three teams stood out by the end of the competition. In the text-to-text category, France’s Wavestone Wavelets worked with Sarvam AI’s Sarvam-30B model.

In the audio-to-text category, India’s MCTE team took the win using Mistral AI’s Voxtral Realtime. And in the image-to-text category, China’s LiteMind team built on Google’s Gemma 4.

Across the board, the winning teams showed that cutting compute needs can come down to model compression and smart engineering choices — and that these optimizations don’t have to come at a major performance cost.

AI built for constrained environments 🌍

Beyond the technical challenge, there’s an energy and economic angle at play too. Models that need less computing power can run in settings where computing and energy resources are limited.

That could help make certain AI technologies more accessible while reducing their footprint on available resources. For Clara Chappaz, France’s Minister Delegate for Artificial Intelligence and Digital Affairs, this push for efficiency addresses several challenges tied to AI development head-on.

« Cutting computational demands makes AI viable even in energy-constrained environments, » she said.

Through this challenge, UNESCO and its partners are backing an approach built around AI that can do more with fewer resources, without giving up on performance.

So, what’s your take: does AI’s future depend on ever-more-powerful models, or on models that can do more with less?