Apple priced the memory, and memory is what a local model actually costs

Apple announced new Mac mini and Mac Studio desktops on August 25, along with two chips: M6, its first 2-nanometer chip, and M5 Ultra, its first quad-die design. There are no new features to speak of — this is a specification bump. What changed is the sales pitch: Apple now describes these machines by how large a model they can hold.

The numbers that decide what fits

M6 goes into the Mac mini with a 12-core CPU, a 12-core GPU with Neural Accelerators, a Dual 16-core Neural Engine and up to 170GB/s of unified memory bandwidth. Memory is 16GB as standard and configurable to 32GB. M5 Ultra goes into the Mac Studio: up to a 36-core CPU, up to an 80-core GPU, 1.2TB/s of unified memory bandwidth — 50 percent more than M3 Ultra — and up to 512GB of unified memory, which Apple says lets users run enormous LLMs entirely on device.

Prices are the other half of the specification. Mac mini with M6 starts at $899 in the U.S. and $799 for education; with M5 Pro it starts at $1,699. Mac Studio with M5 Max starts at $2,499 and with M5 Ultra at $5,499. Both go on pre-order now and ship on September 22.

• Mac mini, M6: 16GB standard, 32GB maximum, 170GB/s, from $899.
• Mac Studio, M5 Max: 18-core CPU, up to 40-core GPU, up to 128GB, from $2,499.
• Mac Studio, M5 Ultra: 36-core CPU, 80-core GPU, 1.2TB/s, up to 512GB, from $5,499.
• Thunderbolt 5 clustering: Apple claims up to 3x faster distributed inference than one system alone.

Why the ceiling matters more than the cores

A model has to fit in memory before speed is a question at all. As a rough working figure — ours, not Apple's — weights at four bits take about half a gigabyte per billion parameters, and about two gigabytes per billion at sixteen bits, before the context window takes its own share. On that arithmetic a 32GB ceiling is a machine for models up to roughly 30 billion parameters at four bits, and 512GB is a machine that holds a 600-billion-parameter model without splitting it.

That is why the two ends of this line are not the same product with different speeds. The cheap one is for small and mid-sized models and for development against them. The expensive one is for running a frontier-sized open model at home or in a small office, which until now meant a rack of specialized GPUs.

The cluster is the part Apple did not invent

Apple's claim of up to 3x faster distributed inference comes from linking several Mac Studio systems over Thunderbolt 5. That path opened in macOS 26.2 last December, whose release notes describe low-latency communication between Thunderbolt 5 hosts for distributed AI inference using MLX, Apple's open framework for its unified-memory architecture. Ars Technica reports that hobbyists and researchers have been daisy-chaining Macs on that path ever since, which is what turned a desktop refresh into an AI product.

Two cautions. Every multiplier here is Apple's own measurement against its own earlier hardware, footnoted rather than published as a method. And a bigger memory ceiling does not make a model good: what a 600-billion-parameter open model is worth against a hosted frontier model is a separate question, and it is not answered by the machine you run it on.