M5 Pro tops M2 Max on LLM speed

The M5 Pro is a sleeper pick for local LLM inference — the tensor API support changes the math significantly compared to prior Apple Silicon generations.

• –Prompt processing (pp512) on the M5 Pro is dramatically faster: 1727 t/s vs 1224 on M2 Max for GPT-OSS 20B MXFP4 (~41% improvement), and 808 vs 554 t/s for Qwen3.5-9B (~46% improvement)
• –Text generation (tg128) is roughly comparable for dense models (30-31 t/s on Qwen 9B), but M5 Pro pulls ahead on MoE: 54 vs 42 t/s for Qwen3.5-35B-A3B
• –The M5 Pro achieves this with 18 GPU cores vs the M2 Max's 38 — the tensor API hardware acceleration more than compensates for the core count disadvantage
• –The M5 Pro has 8GB less RAM than the tested M2 Max (24GB vs 32GB), which limits model selection — the M5 Pro Max with 48-64GB would be a significantly different story
• –M1 Pro users on 16GB are particularly RAM-constrained; the jump to 24GB on even the base M5 Pro opens up MoE models that couldn't fit before