Inference Speed Tests show M5 Max gains

The headline here is not raw token generation, it's prefill. M5 Max looks like a real upgrade for local LLM workflows that spend time on long prompts, summarization, and agentic context loading.

• –Short-prompt generation improves by roughly 14% to 17%, which is solid but not transformative
• –Long-context prompt processing jumps by 2x to 3x, suggesting the M5 Max’s memory and accelerator changes matter most where Macs usually bottleneck
• –That makes the upgrade especially relevant for local RAG, document summarization, and large-context coding assistants
• –The repo is useful because it publishes TTFT, peak memory, and per-run breakdowns, but the results are still community benchmarks, not a controlled lab standard
• –For developers choosing between M4 Max and M5 Max, the M5 looks like a bandwidth and prefill win first, not a universal speed multiplier