Apple announces a Mac Studio with up to 512GB of memory. A sizing calculation explains what capacity tells you, and what still needs a benchmark.
Apple's August 25 announcement puts a larger memory configuration on the Mac Studio roadmap. For someone running local models, the useful question is whether that extra capacity removes a real constraint in their workload. It does not, by itself, answer how quickly the machine will generate a response.
The announcement and its limits
Apple lists up to 512GB of unified memory and 1.2TB/s of memory bandwidth for M5 Ultra. The new Mac Studio is due September 22; the 512GB configuration is scheduled for late October. The US starting price for the M5 Ultra model is $5,499, which is not a quote for the maximum-memory configuration. These are Apple's announced specifications and availability, not measurements from a PeopleAreGeek review unit.
Calculate the weights before the purchase
A useful first approximation is:
Weight storage in bytes = parameter count × bits per weight ÷ 8.
For a hypothetical dense 200-billion-parameter model, that gives 100GB at four bits or 200GB at eight bits, using decimal units. This is arithmetic on a hypothetical model, not a compatibility claim for a particular download. Quantization metadata, runtime buffers, the operating system and the context cache all need additional space.
That distinction changes the buying process. Start with the actual model file and runtime you intend to use. Record its peak memory consumption at your target context length. Repeat with the number of concurrent sessions you need. A single short conversation is a poor proxy for several colleagues querying long documents simultaneously.
Capacity is not throughput
Apple describes the chip architecture separately. Its bandwidth figure is a hardware specification. Dividing that figure by a model's weight size cannot predict end-to-end speed: implementation, memory access, context length and batching affect the result.
For an evaluation, keep four measurements apart: model load time, prompt processing time, time to first token, and generation speed after that first token. Also check output quality with the chosen quantization. A model that fits after aggressive compression may no longer pass the tasks that justified buying the machine.
A decision that can be checked
Compare complete configurations, including memory, storage, tax and the software support you need. Then estimate cost per completed task over your expected usage period, rather than comparing a desktop purchase with a single cloud hour. Include idle periods and maintenance time in that estimate.
The larger memory option is relevant when capacity is the demonstrated bottleneck. If latency or concurrent service is the constraint, wait for measurements of that specific workload. The product photograph above illustrates Apple's announcement; it is not evidence of a test performed here.
September 7 revision: manufacturer sources reviewed, availability clarified and practical analysis rewritten.