AINEWS
The chips in the new Mac mini and Mac Studio mark Apple's first 2nm silicon and first processor built by fusing four chips into one, both engineered for local AI that runs in the background.
Apple reorganized its entire Mac silicon stack on Tuesday around a single bet: that the next AI era runs locally on the device, not in the cloud. The M6 chip in the Mac mini and the new quad-die M5 Ultra in the Mac Studio are not just faster parts. They are Apple's first Mac processors engineered end-to-end for AI workloads that run continuously in the background without round-tripping to a remote server.
The lineup moves in lockstep. The Mac mini now ships with the M6 or the M5 Pro, the first time an M5 Pro has reached a desktop Mac. The Mac Studio pairs the M5 Max with the new M5 Ultra, scaling up the same architectural choices rather than introducing a separate high-end design. Apple tied both product lines to the same on-device AI pitch in its August 25 announcement, framing the chips as headroom for agentic AI, the kind of assistants and automations that work in the background instead of waiting for a user to type a prompt.
Three structural changes carry the pitch. The M6 is the first M-series chip built on a 2-nanometer manufacturing process, the shrink step beyond the "3-nanometer class" label Apple has used for every M5 chip. A 2nm process packs transistors more densely, which generally lets a chip run faster at the same power or cooler at the same speed. The M5 Ultra goes in a different direction: it is the first quad-die M-series chip. Earlier Ultra chips (M1 Ultra, M2 Ultra, M3 Ultra, M4 Ultra) fused two silicon dies into one. The M5 Ultra fuses four, through Apple's UltraFusion interconnect, the high-bandwidth wiring that links separate dies into a single processor the operating system can treat as one.
The M6 also reclassifies Apple's CPU core story. It ships with a 12-core CPU split into three tiers: 2 "super cores" for peak single-threaded work, 4 performance cores, and 6 efficiency cores. Apple has used performance and efficiency cores since the M1; "super cores" are new in the M6, and the layout borrows a page from Apple's phone chips, where a small handful of top-binned cores handle bursty work while the efficiency tier handles the rest. The M6 pairs that CPU with a 12-core GPU that puts a Neural Accelerator on every GPU core, plus a dual 16-core Engine.
The M5 Ultra is the same idea, scaled up. Apple's Mac Studio page lists up to a 36-core CPU, up to an 80-core GPU, a 32-core Neural Engine, up to 512GB of unified memory, and 1.2TB/s of memory bandwidth. Unified memory matters here specifically because the CPU, GPU, and the neural processing unit all draw from the same memory pool, so a large local AI model does not have to be copied between chips to run. The M5 Max, the lower-tier Mac Studio chip, tops out at an 18-core CPU, up to a 40-core GPU, a 16-core Neural Engine, and up to 128GB of unified memory at 614GB/s. The M6 in the Mac mini supports up to 170GB/s of unified memory bandwidth.
All of this lands as Apple-claimed performance, not independent measurement. Apple describes the M6's CPU core as the "world's fastest CPU core" in its press release, and says the M6 Mac mini is "up to 4.8x faster than M4" on its product page. PCMag's explainer lays out the same numbers and explicitly notes that independent benchmarks from Artificial Analysis and the standard test suites are still outstanding. For now, every perf delta above is Apple's claim, not third-party verification.
Two things are worth keeping in view. "Agentic AI on-device" is a real engineering target (local models that can plan, browse, and act) but Apple is shipping hardware headroom for it, not a settled capability. The chips are designed to make the workload possible. The workload, in continuous form, is still mostly a forward-looking pitch. Apple is also not the only company with a local-AI silicon bet. Qualcomm's PC chips, AMD's Ryzen AI 300 series, and Intel's Core Ultra line already pair CPU, GPU, and a dedicated neural processing unit on the same package, with similar unified-memory-style tricks. Apple's move from "3nm class" to a real 2nm node, and the jump from dual-die to quad-die UltraFusion, are the structural deltas. Whether they translate into a real agentic-AI lead is the question the independent benchmarks will have to answer.
Apple says the M6 Mac mini, the M5 Pro Mac mini, and the M5 Max and M5 Ultra Mac Studios all ship in September. The hardware pitch and the AI pitch arrive together. The benchmarks, when they land, will tell the reader whether Apple's bet on local AI was a generation ahead of the field or simply a generation caught up with it.