Local AI on a Mac works because Apple Silicon shares its memory between CPU and GPU: a Mac mini with 32 GB loads a 27B model in 4-bit, a Mac mini M5 Pro with 48 or 64 GB loads it in 8-bit, and a Mac Studio with 256 GB carries models with several hundred billion parameters. As an AI server on your network, the Mac mini runs Ollama or LM Studio, both with an OpenAI-compatible API that any machine in the house can talk to. The limit is memory bandwidth and the Nvidia ecosystem: if you need CUDA, SGLang or NVFP4 checkpoints, you want a DGX Spark.
I've run local models since December 2025 on an Nvidia DGX Spark and on my MacBook Pro M3 Max with 64 GB. I have never owned a Mac mini, and I'll say that up front: my only Mac measurement comes from the MacBook; every Mac mini and Mac Studio figure here comes from Apple's spec sheets, the tools' own docs and ComputerBase's pricing report, each with a date. Everything I write about local AI is collected on my AI page.
This post goes deep on the Mac route. The three-budget comparison is in Your Own AI Server: Hardware and Cost, and the Spark buying guide is DGX Spark: Price and Alternatives.
Why a Mac works for local AI: unified memory and bandwidth
A language model is read from memory in full for every token it generates, so two numbers decide what runs on a machine: how much memory the GPU can address, and how fast it can read it. On a classic graphics card memory is scarce: 24 GB is already the high end. Apple Silicon shares the entire system memory between CPU and GPU, which Apple calls unified memory; a Mac mini with 64 GB hands most of that to the model.
Memory bandwidth is where Macs differ. According to Apple's Mac mini spec sheet, the M6 reads at 153 or 170 GB/s and the M5 Pro at 307 GB/s. The Mac Studio reaches 460 or 614 GB/s with the M5 Max and 1.2 TB/s with the M5 Ultra. Rule of thumb: bandwidth divided by model size in gigabytes is the ceiling for tokens per second. That's a heuristic, not a measurement, but it explains why a Mac Studio runs the same model several times faster than a Mac mini M6.
Mac mini or Mac Studio: which configuration for which model size
Apple refreshed both machines in August 2026. Prices come from ComputerBase's reports of August 25, 2026, because Apple's configurator loads prices via script and can't be cited. The "fits" column is derived from file sizes, not measured.
On pricing: each memory step on the Mac mini M6 costs 220 euros extra according to ComputerBase, and the maxed-out M5 Pro with 64 GB and 8 TB SSD is 7,499 euros. The price of an M5 Pro with 64 GB and a small SSD only shows up in the configurator, so I can't cite it. On the Mac Studio, going from 96 to 256 GB adds 4,400 euros; the 512 GB version is due at the end of October with no published price. The Spark's bandwidth is from Nvidia's DGX Spark product page, as of September 2026.
In short: the Mac mini M6 is for trying things out, not a server for several users. The Mac mini M5 Pro with 48 or 64 GB is what I mean by "Mini as AI server": room for a 27B model in 8-bit, and more bandwidth than the Spark per the spec sheet. The Mac Studio is for models that outgrow the Mini's memory.
Which model fits into how much memory
At 4-bit quantization a model needs roughly 0.5 to 0.6 GB per billion parameters, plus the KV cache for context, which grows with your inputs. That rule of thumb comes from a September 2026 field report on the 512 GB Mac Studio and matches the file sizes I see. A 27B model lands at around 15 to 18 GB in 4-bit and at twice that in 8-bit.
For Qwen 3.8 27B, the model I run on the Spark: Ollama's MLX tag is 18 GB in 4-bit and 56 GB in bf16, and unsloth's 8-bit build is 31 GB. So, without having measured it on a Mac mini: 4-bit fits from 32 GB, 8-bit needs 48 or 64 GB, and the full bf16 version fits in no Mac mini at all.
Going up gets expensive. A 70B class model needs around 35 to 42 GB in 4-bit and runs on a Mac Studio M5 Max with 64 GB; 128 GB leaves headroom for long context. The M5 Ultra field report mentions a 4-bit model of 418 GB that only fits the 512 GB version. Which models are worth considering is in Open Source AI Models Compared.
LM Studio recommends Apple Silicon with macOS 14 and 16 GB; with 8 GB, the docs say, only small models with tight context will run. My take: below 32 GB a Mac isn't fit to be an AI server, because macOS and the context need room next to the model.
My measurement: Qwen 3.8 on the MacBook Pro M3 Max
My only measurement on a Mac, as of August 2026: Qwen 3.8 27B in LM Studio, 8-bit, occupies 27.5 GiB on my MacBook Pro M3 Max with 64 GB and generates 11 tokens per second. Readable, not fluid; for a terminal agent that runs many rounds, it's sluggish. And something no token count shows: on the MacBook, runs broke off more often than on the Spark, and working with the model was a grind overall. That is what drove my buying decision.
For comparison: the same model runs on my DGX Spark as an NVFP4 checkpoint under SGLang at 50.7 tokens per second, measured with 300 tokens in a single stream from the Mac over Tailscale. The two numbers aren't directly comparable: different quantization, runtime and bandwidth. But they show the order of magnitude a 27B model in 8-bit reaches on an M3 Max class Mac. The Spark side is in Qwen 3.8 Locally: DGX Spark Experience.
How fast a Mac mini M5 Pro at 307 GB/s runs the same model, I haven't measured. From the rule of thumb I'd expect somewhat less than 11 tokens per second in 8-bit and more in 4-bit; a hypothesis, not a measurement. I won't be measuring it either: I have never owned a Mac mini and won't buy one for this post.
Setting up a Mac mini as an AI server on your network
A Mac mini as an AI server means the model runs on the Mini, and your MacBook, your Windows machine or a terminal agent talk to it over the network. Both common tools do this out of the box. LM Studio starts a server with an OpenAI-compatible API, default port 1234, reachable at http://<address>:1234/v1. Ollama listens only locally by default; to open it to the network on macOS, the Ollama FAQ says to set the environment variable via launchctl and restart Ollama.
# open Ollama on the Mac mini to other machines on the network
launchctl setenv OLLAMA_HOST "0.0.0.0:11434"
# then quit the Ollama app and start it again
# check from the MacBook that the server responds
curl http://mac-mini.local:11434/api/tags
The second command lists the models on the Mini. The .local hostname works on a home network via Bonjour; I use such addresses in my Raspberry Pi cluster too. For continuous operation, give the Mini a fixed address in your router, turn off sleep, and put a VPN in front of it for access from outside; for my Spark I use Tailscale and never expose a port to the internet. The tools in detail are in Setting Up Local AI: Ollama to SGLang.
On the client side, any program that accepts an OpenAI-compatible address takes the Mini as its endpoint: a chat interface, an editor plugin or a terminal agent like Pi, which I use on the Spark. How that works with a local model is in Claude Code Alternative Without Cloud.
A word about my own AI workforce, so there's no wrong impression: my AI employees run on Claude, not on the Mac. Peter, who does my bookkeeping, is a folder on my computer with personnel file, rules and learnings. When he processed 63 receipts on September 6, 2026, he asked two questions first, one about two currencies, and invented no exchange rates. That way of working lives in the folder, not in the model. A Mac mini with a local model only replaces the model behind it; whether the result is good enough depends on the task.
Mac mini versus DGX Spark: where each one wins
Both are small boxes with shared memory that run a 27B model, yet they solve different problems. The Mac mini M5 Pro wins on entry price, on bandwidth (307 versus 273 GB/s per the vendors' specs) and on everyday use: it's an ordinary Mac, LM Studio installs like any other app, and there's no new operating system to learn.
The Spark wins as soon as the Nvidia ecosystem matters. SGLang, vLLM, NVFP4 checkpoints and speculative decoding, which get me to 50.7 tokens per second, don't exist in that form on the Mac; there you have Ollama, LM Studio and MLX. The Spark has 128 GB in one box while the Mac mini tops out at 64 GB, and Nvidia says several units can be linked for larger models. In return, the Spark is a Linux system that needs care: on mine the services run as systemd user units, the first model start takes around 9 minutes because of the compile step, and one wrong memory parameter freezes the host.
My take: for a small team that wants a 27B model as a chat and writing assistant on its own network and doesn't need CUDA tooling, the Mac mini M5 Pro with 64 GB is the simpler route. For larger models, Nvidia tooling or clusters, take the Spark or the Mac Studio, depending on whether ecosystem or bandwidth matters more. On privacy the same holds for both: the model answers on your own network, and the request to the model needs no cloud provider. What that means legally is in Local AI and Privacy: Cloud or Server?.
Frequently asked questions
Is a Mac mini with 32 GB enough for local AI?
For trying things out, yes: a 27B model in 4-bit takes around 18 GB and leaves room for macOS and context. As a server for several users or a terminal agent it gets tight, because the M6 only has 153 to 170 GB/s. From 48 GB and an M5 Pro upward, the Mini is a usable server.
How much RAM do I need for a 27B model on a Mac?
In 4-bit, roughly 18 GB of model size, so at least 32 GB in the machine. In 8-bit, roughly 31 GB; on my MacBook, Qwen 3.8 27B in LM Studio occupies 27.5 GiB, so 48 or 64 GB. The full bf16 version at 56 GB needs a Mac Studio.
Is a Mac mini cheaper as an AI server than a DGX Spark?
At the entry level, yes: the Mac mini M5 Pro starts at 1,999 euros with 24 GB according to ComputerBase, as of August 2026. With 64 GB and a bigger SSD it approaches Spark prices, which the buying guide breaks down with sources. The Mini is mainly cheaper if 48 GB is enough.
Can I access the Mac mini server from my MacBook?
Yes. LM Studio offers a server on port 1234, and Ollama listens on all addresses once OLLAMA_HOST is set. From the MacBook you reach the Mini via its .local name or fixed IP, from outside the house via a VPN such as Tailscale.
Is a Mac Studio with 512 GB worth it for a small business?
Only if you need a model with several hundred billion parameters locally and know why. For chat, writing help and most agent tasks a 27B model on a Mac mini with 64 GB is enough. The 512 GB version had no published price as of September 2026.
Does an LLM on a Mac also run on a MacBook?
Yes, that's how I measured: Qwen 3.8 27B in 8-bit on a MacBook Pro M3 Max with 64 GB, 11 tokens per second. As a permanent server a MacBook is a poor choice: it gets closed, travels with you and throttles on battery.
Where to go next
If you don't yet know whether it should be a Mac at all, start with Your Own AI Server: Hardware and Cost and its three budgets. Once you've decided, install Ollama or LM Studio following Setting Up Local AI: Ollama to SGLang, pull Qwen 3.8 following Installing Qwen 3.8 with Ollama, and measure how many tokens per second your machine delivers. What local models are good for today, and what they aren't, is in Local AI: What Actually Works in 2026.
