Run Qwen 80B on Mac with 4.3 GB RAM using Swiftlet
August 3, 2026
Swiftlet is a Swift and Metal runtime that enables large MoE models to run on memory-constrained Apple devices by streaming routed weights from storage. It achieves 4.5 to 5 tok/s for Qwen3-Next-80B-A3B on an M5 Mac using only 4.3 GB of resident RAM.
HOW THIS AFFECTS YOU
●
builderYou can deploy large MoE models on edge hardware by utilizing storage-to-RAM streaming.
●
founderThis opens new product opportunities for high-parameter AI capabilities on mobile and consumer Mac hardware.