M1 Max Benchmarks: 72 tok/s Aggregate Throughput at 128k Context

EyalToledano · x · 2026-09-02

Benchmarks for the M1 Max (64GB RAM) show better performance than the M1 Pro due to higher memory bandwidth (400 GB/s). With Q4 quantization, 128k context, and 100% experts in RAM, it achieves 46 tok/s burst and 72 tok/s aggregate throughput with 6 agents. However, memory usage approaches the limit at 256k context.

Original post →

More from Infra

Infra channel →