In a groundbreaking advancement for artificial intelligence in Africa, the continent’s primary multilingual Small Language Model (SLM), known as InkubaLM, has undergone a remarkable 75% reduction in size while maintaining its high performance standards. This significant achievement was made possible through the Buzuzu-Mavi Challenge, an international AI competition organized by Lelapa AI in collaboration with Zindi.
The challenge specifically tasked machine learning experts with compressing InkubaLM, a model tailored to support a variety of African languages, in order to enhance its efficiency and suitability for deployment in resource-constrained settings. Drawing participation from 490 contestants representing 61 different countries, the top positions were clinched by individuals hailing exclusively from Africa — underscoring the continent’s increasing prowess in AI innovation.
Yvan Carré from Cameroon emerged victorious, employing a strategic combination of adapter heads, quantisation techniques, and knowledge distillation to drastically reduce the model’s size. Meanwhile, South Africa’s Stefan Strydom secured second place by implementing vocabulary trimming and shared embeddings to streamline the model to a mere 40 million parameters. The third spot went to the AI_Buzz team comprising Abdourahamane Ide Salifou, Mubarak Muhammad, and Victor Olufemi from Niger and Nigeria; they utilized blended datasets and model distillation methods to create an efficient student model.
This development holds particular significance for Africa where approximately only one-third of the population has consistent access to the internet, with 70% relying on basic smartphones for connectivity. Lean AI models like InkubaLM are designed to operate effectively on such devices even without continuous internet access — empowering practical applications across various sectors including education, agriculture, translation services, and customer support.
“This challenge is not just about technical advancements; it embodies our commitment to developing inclusive AI solutions grounded in African contexts,”
remarked Pelonomi Moiloa, CEO of Lelapa AI.
“When African talent is entrusted with substantial challenges, the outcomes can be truly transformative.”
Celina Lee , CEO of Zindi echoed this sentiment saying: “We’re proud to show that more can be achieved with less. These innovative solutions exemplify what African creators are capable of achieving within the language modeling landscape.”
Looking towards future developments, the most promising models stemming from this competition will play a pivotal role in shaping upcoming iterations of InkubaLM which remains an open-source tool available for adoption by tech communities worldwide. Lelapa AI and Zindi are actively encouraging ongoing collaboration aimed at further refining African AI — making it not only smaller but also smarter and more impactful than ever before.
Leave feedback about this