Running Local LLMs Offline on a Ten-Hour Flight

Dmitri Lerko ran Gemma 4 31B and Qwen 4.6 36B locally during a 10-hour flight with no Wi‑Fi. The MacBook Pro M5 Max had 128GB unified memory and a 40-core GPU; sustained load used about 1% battery per minute, and performance degraded past 100k tokens. The sharp finding is instrumentation: an iPhone cable delivered 60W, while a MacBook cable delivered 94W under the same load.

Why it matters: HKR-H/K/R all pass: this is a named first-person local-inference test with concrete hardware, model, battery, and power numbers. Scope stays practical rather than industry-shaking, so it lands in the 72–77 band.