In an impressive display of ingenuity and skill, Africa’s first multilingual AI model, InkubaLM, has been successfully compressed by 75% while maintaining its performance. This exciting development results from the Buzuzu-Mavi Challenge, a global competition organised by Lelapa AI in partnership with Zindi, aimed at optimising models for low-resource settings.
The challenge attracted 490 participants from 61 countries, all working towards making InkubaLM, which supports multiple African languages, more efficient. The results underscored the growing expertise in AI across Africa, with all top winners hailing from Africa.
Yvan Carré from Cameroon clinched the top spot by combining adapter heads, quantisation, and knowledge distillation. South Africa’s Stefan Strydom followed closely in second place, successfully trimming the model to 40 million parameters through techniques like vocabulary trimming and shared embeddings. The third position was taken by the AI_Buzz team, consisting of Abdourahamane Ide Salifou, Mubarak Muhammad, and Victor Olufemi from Niger and Nigeria, who implemented model distillation along with blended datasets.
The lightweight design of InkubaLM is a significant achievement for Africa, where only 33% of the population has consistent internet access, and about 70% rely on entry-level smartphones. Smaller models like InkubaLM can enhance tools across various sectors, including education, agriculture, translation, and customer service, without constant internet connectivity.
“This isn’t just a technical advancement – it’s a testament to the potential for inclusive, homegrown AI on the continent,” remarked Pelonomi Moiloa, CEO of Lelapa AI.
Zindi CEO Celina Lee added, “These models demonstrate how much can be achieved with less.” The most innovative entries from the challenge will contribute to future open-source iterations of InkubaLM, with both Lelapa AI and Zindi encouraging continued collaboration to push the boundaries of African AI advancements.
