A large model for a small team
One M5 Ultra with 256 or 512 GB runs models that need several cards on a GPU server, for a team where a few people use it at the same time.
Mac Studio for running AI locally
In a Mac Studio the GPU works directly from the machine's unified memory, so a model that fits on no single graphics card runs on one quiet, low-power box on a desk. For larger models we join several Mac Studios over Thunderbolt 5.
Why a Mac for AI
How large a model you can run depends on memory size, and how fast it writes depends mostly on memory speed. A Mac Studio with M5 Ultra has up to 512 GB of unified memory at 1.2 TB/s and draws at most 480 W. We run models on it with MLX, Apple's open-source framework, which most well-known open models are already converted for. The limit is just as clear: reading very long prompts and serving dozens of concurrent users is faster on a GPU server.
What a Mac Studio suits
An office without a server room
A box about 20 by 20 cm, quiet and on ordinary office power; it needs no rack and no separate cooling.
A cluster for trillion-parameter models
Since macOS 26.2, several Macs run one model together over Thunderbolt 5 with RDMA. In Apple's own test, four M5 Ultras ran the 816-billion-parameter Muse Spark Open at 36.7 tokens per second.
Mac mini to start
A Mac mini with M5 Pro has up to 64 GB of memory: enough for small and mid-size models, testing, and an assistant for one department.
Configurations we supply and set up
Specifications from Apple's official pages, reviewed on 22 October 2026.
| Machine | Unified memory | Memory bandwidth | What fits | Suits |
|---|---|---|---|---|
| Mac mini M6 | Up to 32 GB | 153 to 170 GB/s | A model up to about 14B parameters at 8-bit | One person's assistant, testing |
| Mac mini M5 Pro | Up to 64 GB | 307 GB/s | A model of about 30B at 8-bit | A small team with a mid-size model |
| Mac Studio M5 Max | Up to 128 GB | Up to 614 GB/s | A model of about 70B at 8-bit, or an MoE of about 120B at 4-bit | A small team with a large model |
| Mac Studio M5 Ultra | 96, 256 or 512 GB | 1.2 TB/s | With 512 GB, MoE models up to about 700B at 4-bit; DeepSeek-V4-Flash at 8-bit, for example | A very large model on one machine |
| Cluster of two to four Mac Studio M5 Ultra | The machines' memory combined, up to 2 TB | 1.2 TB/s each, joined over Thunderbolt 5 | Models of one to several trillion parameters such as Kimi K2.7 Code or DeepSeek-V4-Pro at 4-bit | Very large models without a GPU server |
| Mac Studio M3 Ultra and M4 Max (previous generation) | M3 Ultra up to 512 GB, M4 Max up to 128 GB | M3 Ultra over 800 GB/s | The same class as M5, more slowly | If you already own one, we set the model up on it |
The "what fits" column is our estimate, not a vendor claim: each billion parameters takes about 1 GB of memory at 8-bit and about half that at 4-bit, and room must be left for context and concurrent users. Apple ships the 512 GB M5 Ultra configuration from late October 2026. A Mac cluster does not merge memory: the model is split across the machines and each runs part of the computation.
The software we set up on a Mac
| Layer | What we install and tune |
|---|---|
| Model serving | MLX and mlx-lm, Apple's open-source framework, or llama.cpp for models with no MLX version |
| Clustering | RDMA over Thunderbolt 5 (macOS 26.2 and later), Apple's JACCL library, and exo to run one model across several machines |
| Programming interface | An OpenAI-compatible API, so your current applications connect to the local model without a rewrite |
| Assistant and documents | A Persian assistant, search over internal documents with citations (RAG), and n8n workflows |
From the first call to a working system
Hardware is chosen after a model has been measured on your own work, not from a table. A quote is issued only once it is clear which model has to serve how many people, and how fast.
- 01
First conversation
We go through the task, the number of users, the data, where the hardware goes and what you already own.
- 02
Measured on your work
We test several open-weight models on samples of your real work and size memory and speed from those results.
- 03
Written quote
You get the exact configuration, the proposed model, the delivery time and the support terms in writing. Prices appear only in the quote.
- 04
Installation
We install the hardware, bring up the model and connect the assistant to your systems, then test it once with international internet cut off.
- 05
Handover and support
You get documentation and training for your team; afterwards we maintain the system and update the models under contract.
How to get in touch
Fill in the form or message us on WhatsApp. The clearer these are, the more exact the quote:
- What the model has to do (answer from documents, code, summarise, process images or speech)
- How many people use it at the same time, and what response time is acceptable
- The model or hardware family you have in mind, if you already know
- Where it goes: your server room, a data centre, or hardware you already own
- How sensitive the data is, and whether it must run with no internet at all
We reply within one business day.
Zharf
Try the ZharfAI assistant right now
Ask about Persian assistants, document automation, or how to start a pilot.
Frequently asked questions about Mac Studio for AI
Is a Mac Studio better than a GPU server?
It depends on how many people use it at once. A Mac Studio gives a lot of memory for less power, and usually less money, than a GPU server with the same memory, and suits a large model with a few concurrent users. An NVIDIA or AMD GPU server reads long prompts faster and suits dozens of concurrent users. We measure this on your real work before recommending either.
How many Mac Studios can be joined together?
In the tests Apple has published, four Mac Studios are joined with a direct Thunderbolt 5 cable between every pair. The cluster does not merge memory: the model is split across the machines and each does part of the work. Apple says four machines are up to about three times faster than one.
What is MLX?
MLX is Apple's open-source framework for running machine-learning models on Apple Silicon. Most well-known open models such as Qwen, Gemma, DeepSeek and GLM have ready MLX versions, and mlx-lm serves them to applications through an OpenAI-compatible API.
Is Persian quality different on a Mac?
No, if the same model runs at the same precision (8-bit, for example). Persian quality depends on the model; the hardware only sets the speed and the largest model you can run. On the Open Persian LLM Leaderboard, Gemma 4 31B and Qwen3.8-27B are among the best open models, and both run on one Mac Studio.
Do you rent GPU servers by the hour or month?
No. ZharfAI does not rent GPUs by the hour or month. There are three ways to work with us: we supply the hardware and install it on your premises, we run a dedicated machine for you and maintain it ourselves, or we set the model up on hardware you already own. If you only need a few hours of compute for a test, say so in the first conversation and we will suggest a suitable route.
How much does it cost?
We do not publish prices, because they depend on the configuration, quantity, the day's market and how it is delivered. After the first conversation and a test on your work you receive a written quote that also names the exact hardware model, the delivery time and the support terms.
Does it work without international internet?
Yes, if the model, the serving software, the assistant, user sign-in and monitoring all sit inside your network. Anything that depends on a foreign service stops during a cut, so before handover we test the system with international internet deliberately cut off.
Tell us which model you want to run
We will say which Mac Studio configuration, or how many machines, it needs, and if a GPU server suits your work better, we will say so plainly.