Mac Studio M5 Ultra: 512GB Makes a Solid Local LLM Case

Apple opened pre-orders on August 25 for a desktop that can hold 512GB of model weights, and the chip inside it is not the new one. The Mac Studio M5 Ultra starts at $5,499 and tops out at 512GB of unified memory, while the newer 2 nm M6 announced the same morning debuts in the Mac mini with a 32GB ceiling. For anyone running models on their own machine, that inversion is the story: Apple’s newest silicon is not the silicon that fits a large model. I have used neither machine, so this is a read of Apple’s own spec sheets against a local setup I do run every week.

The question I brought to the announcement is the one a one-person shop actually faces: does a Mac Studio M5 Ultra replace a cloud subscription, or just move the bill to a different line? My local work goes through Ollama with an open-weight model, mostly for client material I would rather not send anywhere. Everything else runs on paid cloud plans. Three numbers decide whether the Mac Studio M5 Ultra changes that split, and only one of them is the number Apple is leading with.

The Newest Apple Chip Is Not the Local-AI Chip

The M6 is the newer chip and the wrong one for local models. Apple’s chip announcement calls it the company’s first 2 nm chip, with a 12-core CPU complex, a 12-core GPU, and a dual 16-core Neural Engine delivering up to 2x the peak compute of the single engine it replaces. It also caps at 32GB of unified memory and 170GB/s of bandwidth, and it debuts in the Mac mini.

The M5 Ultra is a generation older in name and considerably larger in every dimension that governs whether a model loads at all. Apple describes it as its first quad-die design, with an up-to-36-core CPU, an up-to-80-core GPU, a 32-core Neural Engine, and up to 512GB of unified memory at 1.2TB/s.

Set the two side by side and the buying logic flips:

M6, Mac mini — 2 nm, nearly 30 percent more peak GPU compute for AI than M5, 32GB memory ceiling
M5 Ultra, Mac Studio — older architecture, 16x the memory ceiling, roughly 7x the memory bandwidth

Chip generation is a compute story. Local inference is a capacity story first, and only then a speed story. That ordering is what makes the older part the interesting one.

Mac Studio M5 Ultra Sells Memory, Not Peak Speed

What you buy with a Mac Studio M5 Ultra is headroom, and Apple is unusually direct about it. The Mac Studio announcement says the configuration lets you “run enormous LLMs entirely on device,” and pairs that with 1.2TB/s of memory bandwidth — 50 percent higher than the prior generation — plus up to 4.3x the peak AI compute of M3 Ultra.

The line-up splits cleanly by memory, and so does the price:

M5 Max — from $2,499, up to 128GB of unified memory, up to 3.9x faster AI performance than the prior generation
M5 Ultra — from $5,499, up to 512GB of unified memory, 1.2TB/s bandwidth

Bandwidth deserves a note because it is easy to read as marketing. Token generation on a local model is bound by how fast weights can be pulled from memory, not by how many cores are idle, which is why 1.2TB/s belongs in the same sentence as 512GB. Capacity decides whether the model runs. Bandwidth decides whether waiting for it is tolerable. A Mac Studio M5 Ultra is the first Apple desktop where both numbers point the same direction.

Timing matters if you are budgeting. Pre-orders opened August 25 with general availability on September 22, and Apple notes the 512GB configuration — the entire reason to consider this machine for local work — does not arrive until late October.

32GB Is Where Local Models Stop Getting Interesting

The 32GB ceiling on M6 is the number that decides this for most people, and the arithmetic is plain. A 70-billion-parameter model quantized to four bits is roughly 35GB of weights: parameter count times half a byte, before you allocate a single token of context. That model does not fit in 32GB. At 128GB it fits comfortably with a long context window. At 512GB it fits at full 16-bit precision, around 140GB, with room for a second model resident alongside it.

So the practical tiers look like this, and none of them depend on which chip generation is newest:

32GB — mid-size quantized models, one at a time, short contexts
128GB — a 70B-class model at four bits with real context, comfortably
512GB — large models at higher precision, or several loaded at once

This is where my own experience is worth less than it sounds. What I run locally through Ollama is a mid-size open-weight model for drafts carrying client details I keep off third-party servers, and that workload has never needed anything close to 512GB. When I spent a first week testing an open-weight model on Ollama, the binding constraint was never capacity — it was that I only reach for local inference a few times a week. Buying a Mac Studio M5 Ultra to serve that pattern would be buying a ceiling I never touch.

A memory ceiling only earns its price when your workload keeps hitting it. Mine hits its current one maybe twice a month, and that is the whole answer.

That is the honest version of the local-versus-cloud question, and it is the same trap the last local-model launch to run into hardware math exposed. The specs are real. The frequency is what decides.

$5,499 Is 23 Years of the Cloud Plan I Already Pay For

Put the entry price against a subscription and the gap is absurd on purpose. Claude Pro runs $20 a month billed monthly, per Anthropic’s pricing page, which makes the $5,499 Mac Studio M5 Ultra worth about 274 months of it — roughly 23 years. Configure the 512GB version and the multiple gets worse.

That comparison is unfair, and naming why is the useful part. Hardware is an asset you keep, resell, and use for everything else you do; a subscription is rent on capability you never own. My standing rule is that a paid tier has to return about three times its cost before I keep it, and applied to a Mac Studio M5 Ultra that test stops being about money and starts being about what only local inference can do:

Work that cannot leave the machine for contractual or privacy reasons
Work that has to run with no network at all
Volume high enough that per-token cloud billing outpaces a one-time purchase

The first item is the only one I can speak to from experience, and it is genuinely why I keep a local model at all — the rules I hold for client data on local models exist because some drafts should never touch a third-party API. But that need is measured in a handful of documents a month for me, not in a workload that justifies a $5,499 desktop.

Three Checks Before You Pre-order a Mac Studio M5 Ultra

Three questions answer this faster than any benchmark, and they are all about your last three months rather than Apple’s spec sheet.

Did you hit a memory ceiling? If you have never watched a model fail to load, capacity is not your constraint and the Mac Studio M5 Ultra is solving a problem you do not have.
How many hours a week is the machine actually inferring? Idle silicon is the most expensive kind. Under a few hours, cloud plans win on arithmetic no matter how good the hardware is.
Is any of the work privacy-bound? This is the check that flips the decision, because it is the one cloud pricing cannot compete on at any price.

If two of those land, the M5 Max at $2,499 with 128GB is the configuration to look at first — it clears the 70B-at-four-bits bar that stops the 32GB M6 cold, at less than half the entry price. If all three land, including sustained daily volume, the 512GB Mac Studio M5 Ultra arriving in late October is the machine, and the wait is worth planning around.

For me, this launch is a clean answer to a question I am not asking yet. My local inference is a privacy tool I reach for a few times a week, not a workload, and a 32GB-class machine covers it with margin. What I find genuinely useful is the inversion the announcement makes visible: the newest chip is now the worst one for local models, and anyone shopping on generation number is shopping on the wrong axis. I will revisit this if my local share of client drafting ever gets past a few documents a month — that is the threshold, not the spec sheet.

Sources

AI-assisted research and drafting. Reviewed and published by ToolMint.

In this article

The Newest Apple Chip Is Not the Local-AI Chip

Mac Studio M5 Ultra Sells Memory, Not Peak Speed

32GB Is Where Local Models Stop Getting Interesting

$5,499 Is 23 Years of the Cloud Plan I Already Pay For

Three Checks Before You Pre-order a Mac Studio M5 Ultra

Sources