Swiftlet Runs an 80B Qwen in 4.3GB of RAM on a Mac

Swiftlet exploits the fact that a Qwen mixture-of-experts model activates only about 3B parameters per token. It keeps the dense core resident — 1.3 to 2.5GB at 4-bit — and streams routed experts from SSD, repacked into fixed-stride containers so each fetch is a single read, with LFU-plus-recency cache eviction. Gated DeltaNet linear attention on 75 percent of layers removes the growing KV cache. Reported on an M5 Mac: Qwen3-Next-80B at 4.5 to 5 tokens per second in 4.3GB peak RAM, and a 35B at 7 to 11. The 35B reportedly runs on an iPhone 17 at roughly 1 token per second. Apache 2.0.