Lab

The AIC Lab

We test what we build. The lab is where we run local models, measure them and harden the agent controls that go into our servers.

The AIC Lab rack: two Mac Studio M3 Ultra machines under a status display

The hardware

Two Mac Studios joined by one Thunderbolt 5 cable serve the largest open-weight models as a single server. Models like GLM-5.3 and DeepSeek V4.1 Flash are too big for one Mac, so they run split across both.

The engine is ours: a production-grade inference server we built on Apple's MLX, the same one that runs in the servers we deliver.

Request a quote for this rack →
Machines
2 × Mac Studio with M3 Ultra
Each
32-core CPU, 80-core GPU, 512 GB unified memory
Together
1 TB of unified memory
Link
Thunderbolt 5 with RDMA
Engine
Izuran Pro on our fork of Apple's MLX
Rack
DeskPi RackMate T1

What we work on

MLX and Apple Silicon

Running open-weight models efficiently on Mac hardware, on our own fork of Apple's MLX.

Local inference

One Mac, copies on several Macs, or one model split across two, and what it takes to make that reliable in an office.

Agent harnesses

The controls around an agent: approvals, the network fence around its commands, and the logs.

Evals

Every model and quantization tested on agent tasks before it enters our catalog.

Fine-tuning

Teaching smaller models to use tools well, with adapters for each environment.

The Neural Engine

Izuran's voice and its end-of-turn detection run on Apple's Neural Engine.

Which Mac runs which model

We publish what fits on each size of Mac, with the peak memory and speed we measured, from a 64 GB Mac mini to two Mac Studios.

See the models and Macs →

Where lab Macs go next

When a Mac leaves the lab, we test it, erase it and grade it, and sell it through AIC Lab Store.

Visit AIC Lab Store →

From the lab

Field notes from the lab and from building private AI for real offices.

All Notes →

Want to see it in your office?

Request a trial