Loading page…
The specification refresh formalizes a workflow of linking Macs for larger local models, but performance evidence has yet to arrive and the top chip’s name is inconsistent.
Listen to this story
Apple is using refreshed Mac mini and Mac Studio hardware to make memory capacity and machine-to-machine links the main local-AI selling points. The $899 M6 Mac mini tops out at 32GB, while the $5,499 Studio is described with up to 512GB of unified memory and 1.2TB/s bandwidth—more relevant to running larger models than adding consumer features.
macOS 26.2 adds low-latency Thunderbolt 5 host communication for distributed MLX inference across linked Macs.
Apple claims M6 Mac mini CPU performance up to 40% over M4; independent benchmarks were unavailable at launch.
Both desktops support Wi-Fi 7 and Bluetooth 6; Mac mini includes 2.5Gb Ethernet, with 10Gb optional.
Apple’s new Mac mini and Mac Studio put local AI inference at the center of a desktop refresh that otherwise adds no major features. The bet is on memory, bandwidth and linked-machine workloads rather than a new AI application for the Mac.
Apple is positioning the desktops for local and distributed inference using MLX, an open-source array framework built to use Apple silicon’s unified memory. The approach gives developers a way to run models on their own hardware rather than depend entirely on cloud services.
macOS 26.2 added low-latency communication between Thunderbolt 5 hosts for distributed MLX inference. Developers, researchers and hobbyists have used that capability to chain Mac minis or Studios together, creating enough combined capacity for local language models that would not fit on a single mass-market machine.
The M6 Mac mini starts at $899.
The Mac Studio with M5 Ultra starts at $5,499.
The highest-end Studio configuration is described as supporting up to 512GB of unified memory.
At the lower end, the M6 has a 12-core CPU, 12-core GPU, up to 160GB per second of unified-memory bandwidth and a 32GB memory maximum. Apple says its multi-threaded CPU performance is up to 40% faster than M4, though independent benchmarks were not yet available.
One detail needs clarification. The launch and pricing details call the high-end chip M5 Ultra, while the later hardware description calls it M6 Ultra. The capacity and bandwidth figures are clear, but the conflicting designation leaves the flagship chip’s name unresolved.
Preorders began August 25, and Apple plans to ship both desktops September 22. A 512GB memory configuration is scheduled for late October, leaving early buyers to weigh the published specifications before independent performance testing is available.
Story updates
New reviews from CNET and Engadget portray Apple’s M5 Ultra Mac Studio as an exceptionally fast compact workstation for creative jobs and local AI. They also land on the same constraint: buyers must pay heavily to reach the memory capacities that distinguish the Ultra from cheaper Macs.
The M5 Ultra Mac Studio starts at $5,500, up from $4,000 for the previous Ultra model. Engadget’s test system, with a 36-core CPU, 80-core GPU, 256GB of unified memory and 2TB of storage, cost $11,299. CNET tested a 256GB model with a 4TB SSD priced at about $12,300.
The chip scales to 36 CPU cores, 80 GPU cores and a 32-core Neural Engine, with up to 1.2TB per second of memory bandwidth. Engadget measured a Geekbench 7 multi-core score of 52,185, versus 39,145 for the M3 Ultra Mac Studio, and found gains of roughly 20% to 30% across CPU, GPU and Neural Engine tests.
The Ultra begins with 96GB of memory, then jumps directly to 256GB. CNET says that step costs $4,000, and buyers cannot upgrade the memory later. A 512GB configuration is scheduled for the following month. Apple markets the higher-capacity systems for large local AI models, while the M5 Max tops out at 128GB.
Apple’s local-AI pitch extends beyond one desktop. Up to four Thunderbolt 5-equipped Macs can be clustered to pool memory and processing. Apple says four M5 Ultra systems can process AI work three times faster than one, a claim Engadget did not test. On a single system, however, Engadget transcribed a 70-minute podcast in 18 seconds with MacWhisper—less than one-quarter of the time its tested M4 Max Mac Studio needed.
Engadget found the M5 Ultra only about 10% faster than the M5 Max in synthetic Geekbench AI testing and recommends the Max for most creators and AI developers. The Ultra’s case is strongest when a workload needs more than 128GB of memory, or when compute time is valuable enough to outweigh the price premium.
cnet.comLoading discussion...
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.