Redditor runs Qwen Flash Next locally at 24 tok/s on 64GB MacBook via Strata SSD offloading
New-Pressure-6932 · reddit · 2026-11-07
A user with an M1 Max 64GB MacBook shares a local inference setup: too much RAM for small models, not enough for big ones. Using the open-source Strata project (plus MoEShift, which offloads to SSD), they run Qwen Flash Next locally at 24 tok/s, enough to replace several AI subscriptions for routine tasks. They argue MoE expert offloading deserves more research as a path to running frontier-scale open models on consumer hardware.
More from Infra
- Nvidia requires Southeast Asian cloud partners to enroll in AI chip location tracking program — nmasc_ · 2026-11-07
- Starlink Now Connects 500+ Locomotives for Brazil's Largest Freight Rail Operator Rumo — elonmusk · 2026-11-07
- Tesla targets chip fab inefficiency: wafers spend as little as 20% of time in useful processing — XFreeze · 2026-11-07
- Together Compiler: Agents That Continuously Optimize Kernels, Cutting Latency Up to ~Half — togethercompute · 2026-11-07
- llama.cpp adds Kimi K3.1 Pro support, letting you run Moonshot's giant model locally — jacek2023 · 2026-11-07
- Reactor co-founder publishes 'The real-time manifesto' on serving real-time AI at scale — smtabatabaie · 2026-11-07