WIRED's review of the $11,000 M5 Ultra Mac Studio finds a 122 billion parameter AI model runs locally in under two minutes, while a 15 second video takes 11 minutes.
Apple's Mac Studio is a small silver cube about the size of a coffee mug, and the M5 Ultra version WIRED's Luke Larsen just reviewed costs $11,000 in its tested 256 GB memory configuration. The capability that defines this generation is local: a 122-billion-parameter AI model finished a summarization task in under two minutes on the desktop, with no cloud server in the loop.
"On-device AI" is shorthand for running large AI models on the machine in front of you, instead of renting time on a server elsewhere. Apple leaned on that idea in its August 25 announcement of the M5 Ultra Mac Studio, which adds Neural Accelerators to the GPU and claims up to 4.3x peak AI compute versus the M3 Ultra. Vendor peak-compute figures and an actual workload are different things. The review's value is in the second.
Larsen tested LM Bionic, a local app for running open-weight models, on a 9-billion-parameter Qwen 3.5. Summarizing a dozen local files took roughly two minutes on the Studio. The same task took almost ten minutes on a MacBook Pro with the M5 Max chip, an 18-core CPU, a 40-core GPU, and 48 GB of memory. That gap is real for one reviewer, on one workflow, with no controlled chip-only comparison. It is a working observation about what the machine does on a desk, not a clean hardware win.
A 122-billion-parameter Qwen 3.5 finished the same kind of task in under two minutes, and Larsen described its answers as more accurate and useful than the smaller model's. There is no published quality rubric behind that judgement, so it is a working reviewer's read, not a benchmark. The shape of the box is still clear: a desktop that can hold a foundation-level large language model in memory and answer in the time a person would actually wait.
Video generation told a different story. Larsen asked Draw Things to make a 15-second clip using MiniMax's H3 model, and the local render took more than 11 minutes, with the chassis running hot enough that the fans spun up. The M5 Ultra is fast for text-based local AI; the same on-device promise for high-end video generation is not there yet, and the box makes sure you can hear it working.
The form factor is what most buyers will actually live with. The Mac Studio is the only current Mac with USB-A ports, an SD card slot on the front, and support for up to four 5K displays at 120Hz. The Mac Mini is smaller and cheaper; the Mac Pro is rack-mountable and much more expensive. Apple lists the M5 Ultra Mac Studio starting at $5,499 in the US, with the 512 GB memory configuration coming in late October. The reviewed unit, with 256 GB of unified memory, sits above the shelf model and below the rack-mount Mac Pro.
What that buys is a unit-of-account shift for a narrow professional buyer. Specific high-end AI workloads move from a recurring cloud bill to a single upfront hardware purchase, gated by who can afford an $11,000 desktop and can tolerate the heat and fan noise during sustained AI runs. It is a working tool for the small group of buyers who can already justify a high-end Mac Pro or a rack of rented GPUs, and who want a local loop for the parts of their work that fit on a desk.
The next reference point readers will want is a controlled benchmark. Apple's vendor numbers and a single reviewer's two-model workflow describe the shape of the box, not a clean head-to-head. For now, the M5 Ultra Mac Studio is the rare desktop where "I can run a 122-billion-parameter model at home" is something a buyer can say out loud, and mean it, even if the desk gets loud while they do.