DeepSeek in a backpack: two Mac Studios hit 220 tok/s fully offline

Stefania_druga · x · 2026-11-13

A developer demoed a fully offline, battery-powered portable inference rig: an e-bike battery plus two 256GB M5 Ultra Mac Studios linked via RDMA over Thunderbolt 5 using exo labs, running DeepSeek at 220 tok/sec. He highlighted Apple Silicon's memory density and power efficiency; the demo was billed as the best at the sglang conference.

Related event: Wearable Mac-based compute backpacks steal the show at SGLang conference(2 posts)→

Original post →

More from Infra

Infra channel →