Before you dive in
What you’ll find in the original
- Compare named memory modes on the same suite instead of presenting the fastest configuration alone.
- Define whether peak process RAM excludes the operating system and browser, and distinguish a soft watermark from a hard limit.
- Publish runtime configuration, model revision, device, precision, and limits through an inspectable endpoint.
Worth knowing
On an M5 Pro, the author reports 0.74 GiB at 171 ms in minimal mode and 2.11 GiB at 32 ms in reduced mode. Latency excludes HTTP and uses fixed Pong inputs; JevMade did not reproduce it.