Swiftlet Runs an 80B Qwen in 4.3 GB of RAM on a Mac by Streaming Weights Off the SSD — and a 35B on an iPhone

Show HN·high signalXBlueskyLinkedInCopy link

A Show HN post at 215 points and 98 comments demonstrates Swiftlet, which streams MoE expert weights from SSD instead of resident RAM, fitting an 80B Qwen into 4.3 GB on an M5 Mac at 4.5-5 tokens/sec and a 35B model onto an iPhone. Commenters were split: supporters called it how progress happens, while critics noted prefill is the real bottleneck — roughly half an hour to process 10k tokens on an M5. For builders the interesting part is the architecture bet: MoE's ~3B active parameters per token is what makes disk-backed streaming viable at all, so this technique scales with MoE adoption, not with dense models.