Apple’s New Mac Studio Clusters Local AI on the Desk

Apple put M5 Max and a new M5 Ultra into Mac Studio, with up to 512GB of memory and Thunderbolt 5 clustering for local AI. Pre-orders opened August 25. The first machines arrive September 22, and the 512GB build waits until late October.

The press kit still talks about 8K color and quiet aluminum. The configuration Apple is selling hardest is a desk box that can hold a frontier-class open-weight model, then borrow memory from the machine next to it.

The New Mac Studio Puts a Cluster on the Desk

Mac Studio with M5 Max starts at $2,499 in the U.S., or $2,299 for education. The M5 Ultra model starts at $5,499, or $5,099 for education. Both stay in the same 3.7 by 7.7 by 7.7 inch shell Apple has used since 2022, with the Max at 6.0 pounds and the Ultra at 8.0 pounds.

Johny Srouji, Apple’s chief hardware officer, framed the machine around on-device models rather than a faster timeline in Final Cut.

Mac Studio is the ultimate desktop for on-device AI and the world’s most demanding pro workflows, relied on by users for its tremendous performance and extensive pro connectivity, all in a quiet, compact design that sits right on your desk.

Johny Srouji, chief hardware officer, Apple newsroom

Greg Joswiak, Apple’s senior vice president of marketing, posted the launch video the same night the orders opened.

Mac Studio + M5 Max and the all-new M5 Ultra.

Monumental leap in AI performance. Even faster graphics for the most demanding pro workflows. Massive unified memory and extensive connectivity. All in its compact design. pic.twitter.com/M3iBWZnC5N

— Greg Joswiak (@gregjoz) August 26, 2026

M5 Max uses an 18-core CPU with 6 super cores and 12 performance cores, a 32-core GPU that can be specified as 40 cores, a 16-core Neural Engine, and up to 128GB of unified memory. Base bandwidth is 460GB/s; the 40-core GPU option raises that to 614GB/s. M5 Ultra starts as a 30-core CPU with 10 super cores and 20 performance cores and a 64-core GPU, and it configures to 36 cores (12 super, 24 performance) and an 80-core GPU. Neural Accelerators now live in every GPU core on Ultra as well as Max. Apple describes Ultra as two M5 Max chips fused into one SoC.

512GB of Unified Memory, 1.2TB/s of Bandwidth

The spec that decides whether a local model fits is memory, and the spec that decides whether it talks at a usable pace is bandwidth. M5 Ultra advertises up to 512GB of unified memory and 1.2TB/s of bandwidth, which Apple says is 50 percent higher than the last Ultra. The GPU can address that pool directly, so a large model does not have to shuttle weights across a PCIe link into a smaller card.

Memory that fits the model:

• M5 Max ceiling: 128GB, and only with the 18-core CPU and 40-core GPU option; base machines ship with 36GB and a 512GB SSD.
• M5 Ultra floor: 96GB and a 1TB SSD, with 256GB or 512GB only on the 36-core, 80-core chip.
• Storage cap: 8TB on Max, 16TB on Ultra, on a next-generation SSD Apple rates up to twice as fast as the last Studio, which matters when a model file is hundreds of gigabytes.
• The late SKU: 512GB unified memory is listed, then withheld until late October, so the machine that actually holds the largest weights is not the one that ships on September 22.

Apple’s own tests, run in July 2026 against prior Studios, put M5 Ultra at up to 4.3x the peak AI compute of M3 Ultra and 9.8x M1 Ultra. LM Studio prompt processing is listed at up to 4x M3 Ultra and 9.8x M1 Ultra. Those figures compare a new chip with Neural Accelerators against last year’s Ultra on Apple’s chosen apps. They are not a tokens-per-second score against CUDA.

Thunderbolt 5 Turns Four Boxes Into One Model Host

A single Ultra is the pitch. The second product is a cable. Thunderbolt 5 now carries RDMA, remote direct memory access, so one Studio can read another’s memory without bouncing every byte through the CPU and the network stack. Apple says a cluster of four Mac Studio systems delivers up to 3x faster distributed AI inference than one machine, and that the link builds a shared memory pool for frontier-class open-weight models.

Hobbyists and labs were already doing a rougher version of this. macOS 26.2, which shipped in December 2025, turned on low-latency Thunderbolt 5 traffic for MLX, Apple’s open-source array framework. At WWDC 2026, Apple walked through a four-node M3 Ultra mesh, with mlx.launch and the JACCL library spreading a job across the cables. The new Studio is the first time that setup is a headline SKU rather than a developer-session trick.

How the desk cluster got here

1. December 2025: macOS 26.2 adds RDMA over Thunderbolt 5 for distributed MLX inference between hosts.
2. June 2026: WWDC session on distributed inference shows four M3 Ultra Macs in a Thunderbolt mesh, including a one-trillion-parameter model sharded across the group.
3. August 25, 2026: the new Studio ships six Thunderbolt 5 ports on Ultra (four at the back, two at the front), rates each at up to 120Gb/s, and prints the four-box, 3x inference claim in the product launch.

Four copies of 1.2TB/s of unified memory bandwidth do not become 4.8TB/s of decode in the real world. The interconnect is fast, and it is still a cable. Apple’s 3x claim, not a 4x one, is the number that already admits that. People who have been chaining last year’s Studios say a single fat memory bus still feels cleaner than a model split across boxes, which is why the 512GB SKU, late as it is, is the configuration the local-model crowd is actually arguing about.

What Core AI Changes for Local Models

Hardware without a path into apps is a benchmark machine. Apple is pairing the Studio with Core AI, a macOS 27 framework meant to load and run models on device with a memory-safe Swift API, zero server calls, and zero token invoices. Models get specialized for the chip they land on, with ahead-of-time compilation so the first load is not a long stall. The stack also exposes fine-grained inference memory, zero-copy data paths, and stateful execution, which is the language you use when a 70B-class model has to share RAM with the rest of the desktop.

The on-device toolchain

• Core AI framework: Swift APIs to load, specialize, and run custom models across Mac, iPhone, iPad, and Apple Vision Pro, aimed as the successor to Core ML for this generation.
• PyTorch extensions: export of inference functions into a Core AI artifact, plus hooks for Metal 4 kernels when the default ops are not enough.
• Optimization and debugging: quantization and palettization per layer, an Xcode inspector for graphs, and a macOS debugger that traces runtime data back to the original Python.
• Model recipes: Apple’s GitHub catalog of export recipes for open-source models, with.aimodel files and a Swift package for running them, requiring macOS 27 and Xcode 27.

MLX remains the research path for training and fine-tunes on unified memory. Core AI is the app path, the one that lets a shipping Mac program call Qwen or Mistral without standing up a GPU cloud. Foundation Models on macOS 27 also opened the Language Model protocol so an app can swap Apple’s on-device model, Private Cloud Compute, Core AI, or MLX behind the same session object. That is a software story Apple could have shipped on last year’s Studio. It lands with this one because 512GB and a four-box pool make “bring your own model” a purchase instead of a demo.

A $5,499 Box Versus NVIDIA’s 128GB Spark

The other small box on this desk is NVIDIA’s DGX Spark, a GB10 Grace Blackwell machine with 128 GB of coherent unified memory, 273 GB/s of bandwidth, a 140W chip, and the CUDA stack preinstalled. NVIDIA says one Spark can inference models up to 200 billion parameters and fine-tune up to 70 billion, and that two Sparks linked over ConnectX-7 can handle models up to 405 billion. Retailers have listed Founders Edition units around $4,699 after a February 2026 memory-cost increase. Apple does not run CUDA. NVIDIA does not sell 512GB in a cube that draws from a 480W maximum continuous budget and stays in a 7.7 inch square.

The trade is ugly in a useful way. Spark’s 273 GB/s bus is the reason dense 70B decode feels sticky even when the model fits. Apple’s 1.2TB/s bus is why a 256GB or 512GB Ultra is the box local-model buyers keep putting next to a pair of Sparks at similar money: prefill can be close, decode is not. CUDA still wins when the workload is training, when the tooling expects NIM, and when someone already owns two Sparks and likes the speed they have. People running an M3 Ultra 256GB beside a Spark pair have said the NVIDIA side still feels two to three times quicker on the same open model. The M5 Ultra’s 4.3x AI claim has to beat that feeling in September, not in a slide.

June Already Raised the Last Studio by $1,300

The new prices look like a jump until you look at June. Apple raised the outgoing M4 Max Studio from its March 2025 launch of $1,999 to $2,499, and the M3 Ultra Studio from $3,999 to $5,299 on June 25, 2026, citing memory and storage component costs, as PCMag noted in an editors’ update. The new Max opens at the post-hike Max price. The new Ultra opens $200 above the post-hike Ultra. DRAM has been expensive for everyone in this category; NVIDIA’s own Spark list moved up in February for the same reason.

That math is why the $5,499 Ultra is not the local-AI machine in the brochure. Base Ultra is 96GB. 256GB and 512GB require the 36-core, 80-core chip, and Apple has not printed those upgrade prices in the newsroom note. Storage, the other line that balloons a Studio, still tops out at 16TB. Buyers who opened the configurator on launch day were looking at five-figure totals once memory left 96GB, which is the configuration that can hold the models Apple used as the reason to buy Ultra at all.

Apple Upgrade, a Klarna lease in the U.S., lists $48.99 a month for a 36-month Max lease and $110.10 a month for Ultra, excluding tax and trade-in. That is a payment plan, not a cheaper computer. Memory still cannot be added after purchase, which has been true of every Mac Studio.

Editors Still Get 33 Streams of 8K ProRes

The machine remains a video and 3D box, and those numbers are not small. M5 Ultra’s media engine, with twice the encode and decode blocks of Max, is rated to play up to 33 streams of 8K ProRes 422 at 30 fps at once. Third-generation ray tracing and faster shader cores are the 3D claims. Apple lists up to 3x faster Magic Mask in DaVinci Resolve Studio on M5 Max versus M4 Max, and up to 1.7x faster Redshift scene rendering on M5 Ultra versus M3 Ultra. Ultra drives up to eight 6K displays, or four 5K 120Hz Studio Display XDR units. Max stops at five external displays. Genlock over USB-C can lock a display to a camera, including iPhone 17 Pro.

Ports follow the old split. Both machines put four Thunderbolt 5 ports, two USB-A ports, HDMI 2.1, 10Gb Ethernet, and a headphone jack on the back. Ultra adds two Thunderbolt 5 ports on the front; Max keeps two USB-C ports there. Both keep an SDXC slot. The new piece is Apple’s N1 wireless chip, which brings Wi-Fi 7, Bluetooth 6, and Thread to Studio for the first time. Maximum continuous power is 480W. The enclosure is 35 percent recycled content, including 100 percent recycled aluminum, and Apple says 40 percent of manufacturing energy is renewable.

macOS 27, branded Golden Gate, is in public beta and due as a free update this fall, with Siri AI and the next Apple Intelligence pass bundled in. Those features will land on other Apple silicon Macs. They are not why someone spends Ultra money. The reason is a quiet cube that can hold a model the cloud would meter, and a cable that can pull three more cubes into the same job. The catch on the calendar is blunt enough: the 512GB configuration, the one that makes that pitch literal, does not ship with the September 22 wave. It is due in late October, after the first Max and 96GB Ultra units are already on desks.