A significant advancement in artificial intelligence has been achieved in Africa. The continent's first multilingual small language model, InkubaLM, has been compressed by 75% without sacrificing performance.
This breakthrough makes the AI model more efficient for low-resource environments. The Buzuzu-Mavi Challenge, a global competition hosted by Lelapa AI in partnership with Zindi, facilitated this achievement.
The challenge saw 490 participants from 61 countries working to reduce InkubaLM's size. The model supports various African languages. Winners of the competition were all African, showcasing the continent's growing AI expertise.
Yvan Carré from Cameroon won first place by combining adapter heads, quantization, and knowledge distillation. Stefan Strydom of South Africa placed second, reducing the model to 40 million parameters. The AI_Buzz team from Niger and Nigeria took third place.
InkubaLM's lightweight design is particularly beneficial for Africa. Only 33% of the population has regular internet access, and 70% use entry-level smartphones. Smaller models can power tools in education, agriculture, translation, and customer service.
"This isn't just technical progress - it's proof that inclusive, African-built AI is possible," said Lelapa AI CEO Pelonomi Moiloa. Zindi CEO Celina Lee added, "These models show how much can be done with less." The most promising entries will inform future open-source versions of InkubaLM, with Lelapa AI and Zindi calling for continued collaboration to keep advancing African AI.