Acutis Gate logo Acutis Gate identity for AI at work

A Mac Studio (M5) as the AI server for your office

Acutis Gate · guides for IT directors · updated October 2026

Short answer: a Mac Studio with M5 Max (up to 128 GB) or M5 Ultra (96 to 512 GB) runs large open AI models on its built-in GPU, so one quiet box on a desk can serve a town office, a school or a firm without sending anything to the cloud. Buy memory for the model you want (an M5 Max with 64 or 128 GB runs Qwen 3.8 27B; an M5 Ultra with 256 or 512 GB runs Qwen 3.8 Flash-Next, the 125B model everyone is talking about), run the model in the Ollama app, give staff Open WebUI as their chat, and put an identity-aware gateway in front of your file shares and apps so each person's AI only reaches what that person can. The Mac does not need to join your domain.

Why a Mac Studio

On Apple silicon the GPU uses the same memory as everything else. That is what makes the Mac Studio unusual: it can hold models that would need several server GPUs elsewhere, in one quiet box that sips power. Smaller Macs and CPU-only mini PCs can run small models, but the models that write and reason well enough for real office work need the memory a Mac Studio has.

Which M5, how much memory

  • M5 Max, 64 GB: qwen3.8:27b (an 18 GB download, with tools, vision and thinking). Strong answers for drafting, summaries and questions about your documents.
  • M5 Max, 128 GB: the same qwen3.8:27b with room for many people at once and long documents. (Flash-Next, at 105 GB, does not fit comfortably: macOS lets the GPU use only part of the memory by default.)
  • M5 Ultra, 256 or 512 GB: qwen3.8-flash-next:125b-mlx, Qwen 3.8 Flash-Next in Apple's MLX format (125B parameters, 6B active per word, a 105 GB download), with twice the memory bandwidth of the M5 Max. The big one, for a whole department.

Pick a model that supports tool calling if you want it to work with your systems, not only chat. Leave room: the model plus a few people's conversations should stay under about three quarters of the memory.

Set up the model and the chat

  1. Install the Ollama app from ollama.com/download and open it. It runs the model on the GPU and starts at login.
  2. Install Open WebUI (in Docker Desktop, OrbStack or Colima) and open it in a browser. The first account becomes its admin.
  3. In Open WebUI, Admin Panel → Settings → Models, manage models and pull the one that fits your memory: qwen3.8:27b on an M5 Max, qwen3.8-flash-next:125b-mlx on an M5 Ultra.
  4. Ask it something from your own work and check the speed is comfortable.

Command line

ollama pull qwen3.8:27b                          # M5 Max
ollama pull qwen3.8-flash-next:125b-mlx          # M5 Ultra (105 GB)
ollama run qwen3.8:27b "Summarize what a council work session is in two sentences."

Keep it running like a server

  1. System Settings → Energy: turn on Prevent automatic sleeping when the display is off and Start up automatically after a power failure.
  2. System Settings → General → Login Items: keep Ollama (and your container runtime) there, and set the Mac to sign in automatically, or start the container runtime as a login service.
  3. System Settings → Network → Firewall: allow incoming connections for the container runtime so other PCs can reach the chat over HTTPS.
  4. Give the Mac a fixed IP address (a DHCP reservation) and a DNS name, for example ai.town.example.

Give each person their own access, not one shared account

A chat that sees your file shares through one service account sees everything, for everyone. Acutis Gate installs on the same Mac and sits between the chat and your systems: staff sign in with their Windows logon, their AI opens file shares as them (only the drives Group Policy maps for them, and never the names of folders they cannot open), HR and finance answer by their own permissions, and every action is recorded for your SIEM. We test it on a Mac Studio against an Active Directory domain: the HR person's AI finds the HR files, the finance person's AI never learns they exist.

The Mac does not join the domain: Gate talks to Active Directory with its own service account, and the model keeps running natively on the GPU while Gate, the chat and the HTTPS proxy run in containers. Install it as your normal user, not with sudo:

Command line

curl -fsSLO https://get.acutisgo.com/dl/AcutisGate-Mac.tar.gz
tar xzf AcutisGate-Mac.tar.gz && cd AcutisGate-Linux-*
bash install.sh --name ai.town.example --license ~/acutis-gate.license

The installer uses Docker Desktop, OrbStack or Colima, whichever is running, installs Colima if none is, and picks a model to match the Mac's memory. Then open Gate's Health check, which walks the whole chain and names anything left to fix.

One quiet Mac Studio M5, every person's own reach

Gate installs on the Mac next to the model: Windows sign-in, files as the person, rules and a full audit trail.

Start a 14-day trial Tour the live Gate

Frequently asked questions

Can a Mac Studio be an AI server for an office?

Yes. Apple silicon runs open models on its built-in GPU using the Mac's memory, so a Mac Studio with M5 Max runs Qwen 3.8 27B for a whole office, and an M5 Ultra with 256 GB or more runs Qwen 3.8 Flash-Next (125B), quietly and without sending data to the cloud.

Which Mac Studio runs Qwen 3.8 Flash-Next?

An M5 Ultra with 256 or 512 GB. The MLX build (qwen3.8-flash-next:125b-mlx) is a 105 GB download, which is too tight for a 128 GB M5 Max because macOS lets the GPU use only part of the memory by default. The M5 Max runs Qwen 3.8 27B well.

Does the Mac have to join Active Directory?

Not for Acutis Gate. Gate talks to Active Directory with its own service account from inside its container, signs people in with their Windows logon and opens file shares as each person.

Why a Mac Studio instead of a GPU server?

One Mac Studio holds models that would need several server GPUs, in a quiet box that uses little power and needs no server room. A GPU server is faster per person, at a much higher price, noise and power.

Related guides