Apple puts M6 and M5 Ultra on the desk

Apple has introduced M6 in the Mac mini and M5 Ultra in the Mac Studio. The announcement is less about one magic benchmark than a hardware strategy: put more AI work on the desktop, from everyday assistants to models that need unusually large memory bandwidth.

iyu explainerTwo chips, two jobs

M6 is Apple’s first 2-nanometer chip, with a 12-core CPU, a 12-core GPU, and a dual 16-core Neural Engine. Apple says the new Mac mini can deliver up to four times the AI performance of a prior model, but that number is Apple’s own comparison and depends on workload and configuration.

M5 Ultra takes a different route. Apple says its quad-die architecture reaches up to 36 CPU cores, 80 GPU cores, and 1.2TB/s of unified memory bandwidth. The practical pitch is not a thin laptop experience; it is keeping large creative and AI workloads close to the user.

iyu explainerWhy memory matters

Running a model locally is often limited less by the ability to perform one calculation than by whether the weights and working data fit in memory. More unified memory bandwidth can reduce the wait between those operations, while a larger capacity can keep a bigger model resident.

That does not mean every local model becomes fast or that cloud services become unnecessary. Software optimization, model size, quantization, thermal limits, and the application itself still decide what users feel.

iyu explainerThe useful takeaway

For ordinary users, the Mac mini is positioned as a small always-on computer that can handle productivity, coding, creative work, and selected local AI tasks. For professional users, the Mac Studio targets workloads where memory movement and sustained throughput matter more than a small footprint.

Treat Apple’s “up to” figures as vendor claims to benchmark, not universal results. The durable change is that desktop buyers are being asked to think about local inference as a normal hardware workload.

Treat Apple’s “up to” figures as vendor claims to benchmark, not universal results. The durable change is that desktop buyers are being asked to think about local inference as a normal hardware workload.