Local AI for many agents,
on one Mac.

Most local runtimes are built for one person typing one prompt. ThinkSpread builds Tempo9, a server-style inference runtime for Apple Silicon, and the local apps that run on it.

35B · one stream · tool-call JSON104.9 tok/son one stock 24 GB M5 Pro
80 / 80turns completed in a four-agent soak
547 MBto declare 128K context at startup

Explore Tempo9 See the measurements

104.9 tok/s uses --speculation-k 4 on tool-call JSON; default settings measure 91.8 tok/s. Method and data are published with the results.

The runtime

Tempo9

Agents do not type. They send several requests at once and repeat the same system prompt and tool history on every turn. Tempo9 brings the server-side machinery for that shape of work to a Mac you own: new requests join the batch already in flight, shared prefixes are computed once, KV cache grows with use instead of being reserved up front, and vision work moves to the otherwise idle Neural Engine. Apps connect through an OpenAI, Anthropic or Ollama-compatible endpoint.

Tempo9 1.2.0 is out: brew install thinkspread/tap/tempo9

Overview · Measurements · Manual · Install

Built on Tempo9

Apps that run entirely on your Mac