Apple's new Macs are built for local AI

Apple introduced two new chips on August 25 built specifically to push more AI work onto the desktop: the M6, debuting in a refreshed Mac mini, and the M5 Ultra, debuting in the Mac Studio. Both are aimed at a market Apple has courted less directly until now, developers who want to run and fine-tune large models locally instead of calling an API.

The M6 is a 2-nanometer chip with a 12-core CPU, a 12-core GPU with per-core neural accelerators, and a new dual 16-core Neural Engine that Apple says delivers twice the peak compute of the prior generation. It supports up to 32GB of unified memory at 170GB/s of bandwidth, and Apple says its GPU AI compute jumps nearly 30 percent over the M5, enough for faster prompt processing against on-device language models. The M5 Ultra is a bigger leap: Apple’s first quad-die chip, built by fusing four dies together, with up to a 36-core CPU, an 80-core GPU, a 32-core Neural Engine, and up to 512GB of unified memory moving at 1.2TB/s. Apple says that memory ceiling is enough to run large language models with hundreds of billions of parameters entirely on the machine, no cloud round-trip required. Sri Santhanam, Apple’s VP of Silicon Engineering, described the M6 design goal directly:

M6 combines a new CPU complex, two additional CPU and GPU cores, a Dual 16-core Neural Engine, and more unified memory bandwidth to power through workloads with amazing energy efficiency.

— Sri Santhanam, Apple VP of Silicon Engineering Group

Both machines ship this fall alongside developer-side support in Core ML, Metal, Xcode, and Apple’s Foundation Models framework, and Apple says multiple Mac Studio units can be networked together to run trillion-parameter models split across machines.

What it means for you

Local inference has mostly been a hobbyist or enterprise-datacenter story until now; consumer-grade hardware with a real memory ceiling for genuinely large open-weight models has been the missing piece. A Mac Studio that can hold a few-hundred-billion-parameter model entirely in unified memory changes the calculus for anyone weighing a desktop AI agent against a subscription: you trade a large upfront hardware cost for zero marginal inference cost and no data leaving the machine, which matters if you’re in a regulated field or just tired of API pricing swings. It’s not a fit for every workload, training still belongs on datacenter GPUs, but for running and fine-tuning an open model like the kind covered in on-device frontier models from PrismML and Bonsai, this is the first mainstream consumer hardware built for that job specifically, not repurposed for it.