M1 Max MacBooks throttle under LLMs

This reads less like a dead-end machine and more like sustained-load physics: 35B-class local inference can push Apple silicon into thermal and power limits fast, and Tahoe-era background work may be adding drag. Users who want stable throughput on a Mac need to think about model size, quantization, cooling, and OS activity together.

• –A 64GB M1 Max is capable, but 35B models are still heavy enough to expose thermal headroom and memory-bandwidth ceilings over time.
• –Reports around macOS Tahoe point to higher temps, constant fan use, and background processes like WindowServer or Spotlight, while some users say Sequoia feels cooler.
• –For local LLMs, smaller quantized models usually give better sustained tokens/sec than chasing a large model that initially benchmarks well and then throttles.
• –If the slowdown happens in minutes, it is worth checking fan behavior, ambient temperature, display scaling, login items, and indexing before blaming the chip outright.