Apple's M5 Mac Studio and Mac mini refresh puts 512GB of unified memory on the desk

Apple announced refreshed Mac Studio and Mac mini machines on 25 August 2026, moving both lines to its M5-generation chips and, unusually for a Mac launch, leaning hard on local AI in the marketing. Pre-orders opened the same day, with availability from 22 September. The specifications below are Apple’s own, taken from its newsroom announcement, and the performance multipliers are Apple’s claims against named baselines rather than independent measurements.

What was announced

The Mac Studio comes in two tiers. The M5 Max configuration offers up to a 40-core GPU and up to 128GB of unified memory at a claimed 614GB/s of memory bandwidth, and Apple says it delivers up to 3.9 times faster AI performance than the prior generation. It starts at $2,499. The M5 Ultra configuration is the one that matters most for large-model work: up to an 80-core GPU and up to 512GB of unified memory at a claimed 1.2TB/s, which Apple says is 50 percent higher bandwidth than the previous Ultra. Apple claims up to 4.3 times the peak AI compute of the M3 Ultra and 9.8 times that of the M1 Ultra, and it starts at $5,499. Apple explicitly frames the top machine as able to “run enormous LLMs entirely on device” and to work with “frontier-class open-weight models”, and says multiple units can be clustered over Thunderbolt 5 for distributed inference.

The Mac mini is refreshed too, offered with either an M5 Pro or, at the top, an M6 that Apple describes as its first 2-nanometre chip. It starts at $899. Apple’s positioning here is the low-power, always-on machine rather than the large-model workhorse.

Apple did not publish a list of specific model sizes it tested or the quantisations and context lengths behind its AI-performance figures, so treat the multipliers as vendor claims pending independent testing.

Why it matters for local AI

The single number that changes what is possible is the 512GB memory ceiling. On Apple Silicon, unified memory is shared between the CPU and GPU, so the great majority of it is available to a model. A machine that can hold 512GB in memory can run open-weight models that would otherwise need a multi-GPU server, and it can do so on a single desk at a fraction of the power draw. That is a genuinely different proposition from a discrete graphics card, where 24GB to 32GB is typical on consumer parts and adding capacity means adding whole cards.

The trade-off, which Apple’s framing understandably does not dwell on, is speed. Unified memory buys capacity, but its bandwidth, though high for an integrated design at a claimed 1.2TB/s, still sits below what a high-end discrete GPU delivers to its own dedicated memory, and prompt processing on long inputs in particular tends to be slower on Apple Silicon than on an equivalent-tier NVIDIA card. “Fits in memory” is not the same as “runs fast”. For a large model on the M5 Ultra, expect to be able to run things that would not fit on a consumer GPU at all, at token rates that are usable rather than blistering. As ever, the honest figures will come from independent testing after the machines ship on 22 September.

The Thunderbolt 5 clustering claim is worth watching for the same reason. Splitting a very large model across two or more machines is possible and is already done with tools in the local ecosystem, but it adds interconnect overhead and complexity, and its real-world throughput on these machines is not something we can confirm from an announcement.

The history

This launch is, in effect, a reversal of the story we covered in May, when Apple narrowed the high-memory Mac Studio options during a DRAM squeeze. The very high-capacity configurations that made the machine a favourite for large models were trimmed then; the M5 Ultra now restores and extends the top of that range to 512GB. If you held off a purchase because the configuration you wanted had disappeared, this is the update to look at, though the same advice applies as always: confirm the exact configuration and current price before committing.

The wider context is a market in which unified memory has become a recognised route to running large models locally. AMD’s Strix Halo platform put large unified-memory allocations on x86 machines earlier this year, and software support has been keeping pace, with Ollama and others tuning specifically for Apple Silicon. Apple pushing its memory ceiling back up, and marketing it explicitly for local AI, is a notable signal that it sees this audience as worth competing for.

The software side

Getting the most from these machines locally means using software built for Apple Silicon. MLX, Apple’s own array framework, and MLX-LM are the native path and tend to make the best use of unified memory and the GPU; llama.cpp and Ollama both support Metal and are the common general-purpose routes. Model availability in the MLX format has grown considerably, which matters because a machine like this is only as useful as the models you can actually load on it. The relevant hardware number to watch alongside memory capacity is memory bandwidth: with a very large model held in memory, how fast the machine can read those weights each token is a large part of what determines how quickly it responds.

Worth it if

You want to run large open-weight models locally, value capacity and low power over outright speed, and would otherwise be looking at a multi-GPU build or a rented server. The M5 Ultra with a large memory configuration is one of the few single desktop machines that can hold a very large model at all. If your priority is fast inference on models that already fit comfortably on a 24GB to 32GB card, a discrete GPU remains the better value, and our cost calculator and machine speccer are the place to check that against your own workload before spending.