Apple launched the new Mac mini on August 25, and for once the AI part is not hiding in a footnote. The base machine now gets the new M6 chip, Apple talks about local models and desktop agents right on the product page, and LM Studio is used in Apple’s own performance tests. The little silver box is being sold as an AI computer now.
Then I reached the memory line.
The $899 M6 Mac mini starts with 16GB of unified memory and can be configured only up to 32GB. Apple says the same chip can deliver up to 4 times faster AI performance than the M4 in its tests, and memory bandwidth rises as high as 170GB/s. Nice. But the more expensive Mac mini uses an older M5 Pro chip, not an M6 Pro, and that machine goes to 64GB with 307GB/s of bandwidth. For local LLM work, I think this is the part buyers should care about more than the M6 badge.
Access without medium partner: Apple Mac Mini M6 Model Details
The M6 upgrade is much bigger than a normal Mac mini refresh
The new Mac mini keeps the same five inch square design Apple introduced with the M4 generation. Two USB C ports and a headphone jack stay on the front. HDMI, Ethernet and three Thunderbolt ports sit on the back. The M6 version uses Thunderbolt 4, while the M5 Pro model gets Thunderbolt 5. Apple also moved standard Ethernet from 1Gb to 2.5Gb and added Wi Fi 7 plus Bluetooth 6 through its N1 networking chip.
Most of the change is inside.
The M6 has a 12 core CPU and 12 core GPU. Apple added Neural Accelerators to every GPU core and built a Dual 16 core Neural Engine into the chip. The M6 also pushes memory bandwidth to as much as 170GB/s. Apple’s previous M4 Mac mini had a 10 core CPU, 10 core GPU and 120GB/s of memory bandwidth in the configuration Apple compares against.
Apple’s own LM Studio result is the number that caught my eye. The company says the M6 Mac mini can reach up to 13.5 times the time to first token performance of the old M1 Mac mini and 4.8 times the M4 in its test. Apple ran these tests in July 2026, so this is fresh data rather than a number carried over from last year.
There is a catch inside that sentence. Apple is talking about prompt processing and time to first token, not telling us that every local model will generate text 4.8 times faster. Those are different parts of inference. A machine can chew through a prompt quickly and then write the answer at a much lower rate once memory traffic starts to dominate. We already see this on other unified memory AI systems.
I spent more time than I expected checking Apple’s wording because the product page and newsroom use slightly different comparisons. This stuff gets messy fast. The safest version is simple: M6 has much more AI hardware than M4, and Apple has measured big gains in selected LM Studio work. We still need independent tests after the machine ships on September 22.
That last date matters. Nobody outside Apple’s test group has a retail M6 Mac mini sitting on a desk yet.
32GB is where the local AI story gets awkward
Apple has spent the last two years proving why unified memory is useful for local AI. The CPU and GPU can work from the same pool instead of forcing a model to fit inside a separate VRAM budget. A 64GB Mac can load models that simply do not belong on a 24GB gaming GPU without heavier compression or offload.
So 32GB feels odd on a new AI focused Mac mini.
A rough four bit weight estimate makes the problem easy to see. A 27B model needs around 13.5GB just for weights. A 32B model is around 16GB. A 70B dense model starts near 35GB before runtime overhead, context cache and macOS take their share. The exact files vary because quant formats add metadata and models are built differently, but the direction is clear.
The M6 Mac mini should be a nice home for 8B, 14B, 20B and 27B class models. Some 32B setups should also be comfortable on a 32GB machine if you keep context and other memory use sensible. A dense 70B model is a different job. You are already beyond the full 32GB pool before you give macOS a chair at the table.
And macOS needs memory too.
This sounds obvious, but local AI buyers keep getting trapped by installed memory numbers. You cannot give every byte of a 32GB Mac to the model. The operating system stays alive. LM Studio or MLX needs memory. Your browser probably has twelve tabs open because we are all terrible at closing them. Long context can eat several more gigabytes. A local agent may also run embeddings, search, code tools or a second small model.
Thirty two gigabytes can go from “that is plenty” to “why did this model just fail to load?” very quickly.
This is why I would not buy the base 16GB M6 Mac mini mainly for local LLMs. It can run local AI, yes. Small models have become very good. Apple’s own AI features will work. But if local models are the main reason for purchase, 16GB gives you a short runway.
The 32GB version is the one that makes sense to me.
Still, Apple stops there.
The older M5 Pro is the better local AI chip in some ways
This new Mac mini lineup looks backwards at first.
The cheaper machine gets Apple’s newest M6. The $1,699 model gets M5 Pro, a chip from the previous generation. Apple did not announce an M6 Pro. So a buyer can spend more money and get an older generation number printed on the silicon.
For normal shopping, that sounds like the sort of thing somebody in a shop has to explain twice.
For local AI, the M5 Pro option makes sense.
Apple lets the M5 Pro Mac mini reach 64GB of unified memory, twice the M6 ceiling. Memory bandwidth is 307GB/s instead of 170GB/s. It can also be configured with up to an 18 core CPU and 20 core GPU. Thunderbolt 5 is included, which matters if you plan to connect fast storage or start doing weird Mac mini cluster experiments.
Apple itself now mentions clustering Mac mini systems on the product page. It names exo and LM Studio Bionic as examples for running multiple AI agents across several machines. I did not expect Apple to put that sort of thing on a consumer Mac page a few years ago. Now it is part of the sales pitch.
For a large local model, 64GB changes what you can load. A 70B model around four bit becomes realistic. You have room left for macOS and cache. Smaller models can run with much longer context or alongside other local services. The extra bandwidth also helps generation because large LLM decode can become a memory traffic problem.
So yes, the newer M6 has newer AI blocks.
I would still choose the M5 Pro Mac mini over it for a workload that needs 40GB or 50GB of model memory. The M6 cannot compute on weights that do not fit.
That is the weird bit.
Apple skipping M6 Pro explains the strange lineup
Apple has not given a detailed public explanation for why the Mac mini jumps from M6 to M5 Pro. A recent report from MacRumors, citing Bloomberg’s Mark Gurman, says Apple is skipping M6 Pro and M6 Max and moving faster toward M7 because the M7 family is expected to carry bigger changes for on device AI and GPU work.
Treat that as reporting, not an Apple promise. Apple has confirmed the product in front of us. It has not announced the full M7 roadmap.
But the shipping lineup itself tells us enough. There is no M6 Pro Mac mini in August 2026. The higher tier uses M5 Pro. MacRumors also reported on August 28 that the base M6 may have a fairly short life before M7 arrives in the first half of 2027.
I normally dislike the “wait for the next chip” game because there is always another chip. You can wait yourself into never buying anything. This case is a bit different because Apple has already made the M6 generation unusual. The new chip launched without the normal Pro and Max family around it.
If I needed a Mac mini today, I would not care. Well, today means preorder today, because shipping starts September 22.
If I were spending a lot mainly for local AI and could wait six months, I would at least watch what Apple does with M7.
The annoying part is that M6 looks like both a big step and a bridge at the same time.
The $899 starting price changes the old Mac mini argument
Mac mini used to have an easy superpower: it was cheap for an Apple computer.
The M4 model launched in 2024 at $599 in the United States. Apple later raised that entry price to $799 in June 2026. The new M6 version now begins at $899 with 16GB memory and 256GB storage. That is a 50 percent jump from the original M4 starting price.
Macworld put it pretty plainly this week: this is no longer a budget Mac.
I agree.
Nine hundred dollars for an M6 desktop is not outrageous when you look at the CPU, GPU, tiny case, macOS and power use. But the old “just buy the $599 Mac mini” recommendation is dead. Once you start adding memory and storage for local AI, the price climbs into a very different conversation.
And storage deserves a mention. The base system is still 256GB. Apple says storage performance is up to twice as fast this generation, but 256GB is tiny if you plan to keep local models around. A few quantized LLMs, image models, caches and Xcode can make that SSD feel like a hotel room during wedding season.
External storage is easy enough on a desktop. I would rather spend on memory first because you can add a fast external SSD later. You cannot upgrade unified memory after buying the machine.
That is the configuration trap I would watch.
Apple gives the M6 a much better chip, then pairs the entry model with 16GB and 256GB. The machine starts looking far more interesting for local AI after you move away from the starting configuration that gets the headline price.
M6 is probably better for small models than the spec sheet makes it sound
I have spent most of this article complaining about 32GB, so I should correct the mood a bit.
Small models in 2026 are very good.
You do not need a 70B model for every job. A local coding helper, document search tool, voice pipeline or private writing assistant can work with models far below that size. Mixture of Experts designs make parameter count even less useful as a simple speed guide. Some models carry a large total weight set but activate a much smaller portion for each token.
For those jobs, M6 looks interesting.
The chip has 170GB/s bandwidth at the top configuration, Neural Accelerators in every GPU core and Apple’s new dual Neural Engine. MLX and LM Studio already make Apple Silicon much easier to use than it was during the M1 days. The computer is tiny, quiet, and can sit on a network as an always on local AI box without a 500W graphics card heating the room.
Apple is leaning into that exact use. Its August launch material repeatedly talks about always on agents and local models. The product page even shows Perplexity and calls out agentic workflows.
That is where I think the $899 Mac mini could find a new kind of buyer. Not somebody trying to cram the biggest possible model into it. Somebody who wants a small model server that also happens to be a normal Mac.
There is something funny about that. The Mac mini spent years as the cheap desktop people put behind televisions, inside offices and under developer monitors. Now Apple is telling people to use a stack of them for agents.
Fair enough.
M6 versus M5 Pro is really capacity versus newer silicon
If I were buying the new Mac mini for normal desktop work, I would start with M6. The CPU is newer, the GPU has the new accelerator setup, and most people do not need 64GB. Paying $800 more just to move to M5 Pro is hard to justify for email, web work, light code and a few local AI tasks.
For local AI as the main job, my choice changes.
I would take 64GB and 307GB/s before I took a newer chip capped at 32GB. A model that fits badly is a bad model experience. Memory bandwidth also matters once generation starts. The M5 Pro is older on the calendar, but its memory setup is much closer to what I want from a serious local LLM machine.
If 64GB is still not enough, I would stop configuring Mac mini and look at Mac Studio. Apple’s new Studio now goes far beyond the mini in memory capacity. At some point, adding more small boxes becomes a fun project instead of the sensible purchase.
There is also a nice middle case. A 32GB M6 Mac mini with a good 14B to 27B model could be the better daily machine than a 64GB system running a huge model slowly just because it can.
Bigger is not always better here.
I know, after spending half this article asking for more memory.
The M6 Mac mini is good. The name is slightly distracting
The M6 Mac mini is one of the freshest computers you can preorder right now. It has Apple’s first M6 chip, a much faster GPU setup, Wi Fi 7, standard 2.5Gb Ethernet, faster storage and a clear focus on local AI. It arrives September 22 with macOS 27.
But I would ignore the M6 label for five minutes before ordering.
Look at memory first.
If your local AI work fits comfortably inside 24GB or 32GB, the M6 Mac mini could be a very neat little box. If you want dense 70B models, several local agents at once, big context or more room to experiment, the older M5 Pro model is the one that starts making sense despite the higher price and older generation name.
That is probably the strangest thing about Apple’s new Mac mini.
The newest chip is not automatically the better AI computer.
And for a machine Apple now openly sells around local AI, the 32GB ceiling is the spec I would remember after all the launch slides are gone.