Back in November, Apple quietly published a research article about the Neural Accelerators in the M5 chip. The numbers are wild.

The base M5 MacBook Pro already delivers up to 4x faster time-to-first-token compared to the M4 when running large language models through MLX. Image generation with FLUX is 3.8x faster. This is on the base chip with 24GB of unified memory.

Think about what happens when the M5 Pro and M5 Max show up with more memory bandwidth and more Neural Accelerators. And eventually the M5 Ultra in the Mac Studio.

Right now, people serious about running local AI often look at expensive PC builds with dedicated GPUs. The M5 generation might change that math entirely. A well-configured M5 Max MacBook Pro or Mac Studio could become the machine for people who want to run models locally, privately, on their own hardware.

Apple’s unified memory architecture was always a theoretical advantage for AI workloads. With the M5’s Neural Accelerators, that advantage is becoming very real. If you’re interested in local AI and you’re on an M3 or earlier, I’d wait for these announcements before buying anything.