AI HOT (Curated Pool)Aug 25, 2026, 21:00 UTC+8

Apple launches Mac Studio with M5 Max and M5 Ultra, built for on-device LLMs

Apple 推出搭载 M5 Max 与 M5 Ultra 的全新 Mac Studio

The new Mac Studio packs M5 Max and M5 Ultra with Neural Accelerators inside each GPU core, delivering up to 4.3x faster AI performance. The M5 Ultra config supports up to 512GB unified memory and 1.2TB/s bandwidth, letting users run frontier LLMs entirely on-device. Four units can cluster over Thunderbolt 5 for 3x faster distributed inference. Pre-orders start today, available September 22.

Why it matters: Official Apple launch of M5 Max/Ultra Mac Studio with up to 4.3x AI perf gain and 512GB unified memory — lowers the bar for local LLM inference. Score stays at featured threshold because this is a hardware refresh, not a model/capability breakthrough, and the source is a press...

Read the original ↗Export Markdown

Take

The number that matters here is 512GB unified memory at 1.2TB/s bandwidth. Until now, running a 400B+ model locally meant either multiple H100s or painfully slow CPU offloading. Apple's approach puts the memory right next to the chip, sidestepping inter-GPU communication overhead entirely.

I'd discount the marketing a bit. Apple says 'frontier LLMs' but doesn't name which ones. The 4.3x AI performance claim is versus M3 Ultra, not Nvidia. And the four-unit Thunderbolt 5 cluster sounds neat, but the post only says '3x faster than a single system' — no actual throughput numbers.

The real use case: AI researchers prototyping models, data scientists running local inference without per-token billing or cluster queues. If you're just running a 70B model, the M5 Max tier probably does the job. The Ultra premium only makes sense if your model genuinely needs that memory headroom.

Source: AI HOT (Curated Pool) · apple.com
Published: Aug 25, 2026, 21:00 UTC+8 · Aug 25, 2026