agi.jetzt
DEEN

What runs our local measurements?

Measuring a model locally means it computes on a machine that stands with us, not in a data centre. This page takes that machine apart layer by layer, from the housing to the program that starts a run and records it.

Not every run computes here

  • Self-hosted (local)The model weights are on this machine, and the computing happens here.This page
  • Managed APIThe model computes at the provider. Our side sends the task and records the answer.Not this machine

Every published run names its runtime.

Schematic exploded view

  • Metal: hardware
  • Glass: software

  1. 01 / 09Hardware

    Mac Studio

    The machine for local runs: a Mac Studio set aside for measuring. Whatever is measured locally computes inside this housing.

  2. 02 / 09Hardware

    Under the housing

    From top to bottom: the aluminium housing, the cooling with fan and heat sink, the logic board, and below it the internal frame, the power supply and the bottom cover. The drawing is schematic: the order of the parts follows Apple’s repair documentation, shapes and sizes are simplified.

  3. 03 / 09Hardware

    Apple M4 Max

    The chip on the logic board. It combines CPU, GPU and Neural Engine in one component. This is where a locally run model computes.

  4. 04 / 09Hardware

    Unified memory

    64 GB

    The memory is not a plug-in module. It sits in one component with the processor, and CPU and GPU use the same memory. It decides how large a model can run locally: the weights have to fit.

  5. 05 / 09Hardware

    Local storage

    1 TB SSD

    The SSD module holds the model weights, caches and the files a run produces. Unlike the memory, it is a part that plugs into the logic board.

  6. 06 / 09Software

    macOS and Metal

    macOS 27

    From here on there is nothing to touch: the upper layers are software, which is why they are drawn as glass. macOS manages the machine. Metal is the interface through which programs compute on the GPU.

  7. 07 / 09Software

    Local inference

    A runtime loads the weights into memory and lets the model compute, for example MLX, llama.cpp or Ollama. They stand side by side, not on top of each other: a run uses one of them and names it.

  8. 08 / 09Software

    Prüfstand runner

    agi.jetzt’s own program that carries out a run: load the task, start the model, record the result, check it, and write down what it was measured with.

  9. 09 / 09

    The whole rig

    Hardware, the runtime on it, the runner on top: a local measurement comes out of these three layers. That is why a result always describes the model in this setup, not the model as such.

Configuration

Machine
Mac Studio
Chip
Apple M4 Max
Unified memory
64 GB
Internal storage
1 TB SSD
Operating system
macOS 27

As confirmed on 10 Oct 2026. Core counts, power draw and energy per run are not recorded. The runs of the pilot name their runtime, not a machine.

← Back to the Prüfstand