FaaSMoE: Serverless MoE Serving Uses <1/3 Resources via Expert-as-FaaS Decoupling (EdgeSys 2026)

arXiv (EdgeSys 2026)·medium signal

FaaSMoE, accepted at EdgeSys 2026, presents a serverless architecture that deploys MoE model experts as stateless FaaS functions, decoupling control and execution planes for multi-tenant serving. Tested on Qwen1.5-MoE-2.7B, the system uses less than one-third of baseline resources while enabling scale-to-zero capability and configurable expert granularity. For agent workloads running MoE models (increasingly the default architecture), this pattern could dramatically reduce inference costs in multi-tenant agent platforms.