M5 Max owners test on-battery inference

If you care about running local models unplugged, this is the right question: sustained tok/s per watt matters more than peak throughput.

• –The comparison is really about usable performance while capped by battery and thermal limits, not charger-attached benchmark numbers.
• –Model size, quantization, and runtime choice will move the needle a lot; MLX, Ollama, and how well the stack uses Apple’s accelerators all matter.
• –If M5 Max stays responsive on battery without collapsing under throttling, it becomes a stronger portable inference platform than many x86/AMD laptops.
• –The thread reflects where local AI hardware buying is headed: developers now judge laptops by unplugged inference behavior, not just specs on paper.