Everything on the launch page comes from a measurement you can rerun yourself. The harness, the raw memtier output and every flag we set are in the public repo: https://github.com/hotkv-ltd/hotkv
I also recorded two sessions on a single 160 vCPU Azure box, same memtier line, one server after the other, with the load generator on the same machine (so the absolute numbers are lower than the published run):
A multi-threaded, Redis-compatible in-memory data store written in C++20. Existing clients connect unchanged. Vector search, a semantic cache for LLM response and agent memory run inside the server. Benchmarks with Redis, Valkey and Dragonfly are public.
