mixture of experts

Researchers developed an AI model that achieves near-full performance using only 12.5% of its expert modules in a mixture-of-experts architecture. This efficiency breakthrough could significantly reduce computational costs and energy consumption for large-scale AI deployments.