LOCAL AIMODEL / INFERENCE / HARNESS
Three layers.
One working setup.
Application
& harnessPi
The working session, context, and tools.
Chosen for its lean harness and minimal system prompt. Requests & responses · compatible API
InferenceoMLX
Loads the model, runs inference, and serves the response.
Optimized for Apple Silicon. Model execution
ModelQwen3.8-27B
Generates responses from the context it receives.
A configuration that fits the available memory. MacBook Pro M5 Pro48 GB unified memory
LOCAL AI SYSTEMSCurrent setup
Local AI
Built around
the hardware I have.
I run Qwen3.8-27B through oMLX, with Pi as the agent harness, on a MacBook Pro with M5 Pro and 48 GB of unified memory. I chose the pieces together: a model configuration that fits, inference optimized for Apple Silicon, and a lightweight harness.
THE DESIGN GOALGet the most useful performance from the machine by matching the model to memory and keeping the surrounding overhead low.