Apple Mac Studio (M3 Ultra)

The Mac Studio with the M3 Ultra is the machine you buy when nothing else will hold your model. 512GB of unified memory, fed at 819 GB/s - roughly triple the bandwidth of the Strix Halo and DGX boxes - in a silent desktop that draws less power than a games console. It’s the closest thing to a personal AI workstation that just works, and it’s priced to match.

The take

This is the gold standard for loading big, and the numbers back it up. 512GB means you can run 400B-class models locally that no other desk-side box comes near, and the 819 GB/s of bandwidth is the fastest of any machine here by a distance, so once a model fits, it answers quicker than the PC-based rivals could dream of. The trade is money and prompt processing. It’s around ten grand kitted out like this, and while token generation is quick, feeding it a very long prompt takes its time - owners running huge models notice the wait before the first token. Weigh that carefully before the money leaves your account.

What it’ll actually run

This is where the 512GB does something no rival can. A 70B at Q4 is trivial, it barely notices, and you can climb all the way to 400B-plus mixture-of-experts models like the full DeepSeek R1, which people really do run on this machine at home. Load a dense 70B at high precision, keep several big models resident at once, or hand a model a huge context window: the capacity is simply in a different league.

Speed splits two ways. Token generation rides that 819 GB/s and is quick even on large models. Prompt processing - the bit where it reads your input before replying - is the M3 Ultra’s soft spot, and on a very long prompt or a massive model you’ll wait for the first token. MLX is the stack to use, it’s matured nicely on Apple silicon, and the machine stays cool and near silent through all of it. The usual Apple caveat applies: no CUDA, macOS only, so Nvidia-specific tooling is out.

Who should buy it

Buy it if you need to run models that won’t fit anywhere else, and you’re happy on a Mac. For a researcher or developer prototyping frontier-size models at home, it’s the one box that does the job in silence, on your desk, without a cloud meter running.

Give it a miss if value is the deciding factor, or if your models fit in 128GB. A Strix Halo box loads a 70B for a fifth of the price, and if you don’t need more than 128GB you’re paying thousands for capacity you’ll never touch. But for the very biggest models, run locally and quietly, this is the machine. There isn’t really a second place.