Tiny but mighty: This AI mini-model outsmarted Microsoft on Meta’s GAIA benchmark

Coral Protocol, a London-based AI company, has achieved a significant milestone by developing a multi-agent AI "mini-model" system that outperformed Microsoft’s agent platform by approximately 34% on Meta’s GAIA benchmark. GAIA is a challenging test suite comprising nearly 450 complex real-world tasks requiring reasoning, web browsing, data analysis, and tool use. While human participants typically answer about 92% of GAIA questions correctly, advanced large models like GPT-4 manage only around 15%. Coral’s mini-model scored the highest among small-scale AI systems, surpassing Microsoft-backed Magnetic-UI, which scored about 30%. Coral’s approach diverges from the traditional AI scaling method of building massive models with billions of parameters. Instead, it employs horizontal scaling by orchestrating many specialized, lightweight mini-models that collaborate in real time, each excelling at specific tasks such as natural language understanding or coding. This collective intelligence framework enables faster, more cost-effective, and potentially more secure